Patent Recommendation Methods for Heterogeneous Enterprise Technology Demands in the Lithium Battery Industry

Abstract

Patents are essential carriers of technological innovation, and their efficient transfer is critical for accelerating technological iteration in the lithium battery industry and supporting sustainability in the new energy sector. However, existing patent recommendation methods lack frameworks for handling heterogeneous enterprise demands, which limits the accuracy of supply–demand matching. This study proposes a knowledge graph-based differentiated patent recommendation framework for enterprise technological demands in the lithium battery domain. A five-element content framework—material, method, efficacy, product, and application—is constructed from both the supply and demand sides. Enterprise demands are classified into complete and incomplete types based on element coverage, and patent supply knowledge graphs are built for potentially relevant patents. Two differentiated recommendation methods are then developed. For complete demands, the Precision Recommendation Method for Complete Technological Demands integrates BERT-based semantic encoding, TransE-based structural modeling, and RAG-based constraint retrieval to achieve precise matching under full element coverage. For incomplete demands, the Fuzzy Recommendation Method for Incomplete Technological Demands incorporates multi-source enterprise data to enrich demand categories and constructs augmented query contexts to generate diversified candidate patent sets. Empirical validation based on 25 demand-driven patent transfer cases shows that the PR-CTD method exactly identifies the actual transferred patents in three cases. The FR-ITD method ranks 6 out of 14 actual transferred patents within the Top-5 results, while the remaining cases are all within the Top 13. These results demonstrate the effectiveness of the proposed framework in real-world patent transfer scenarios. This study provides a novel theoretical perspective for the structured modeling of heterogeneous technological demands and supply–demand semantic matching. It also offers practical value by improving the efficiency of patent retrieval and matching, thereby supporting patent technology transfer in the lithium battery industry.

1. Introduction

As the core power source for new energy vehicles, large-scale energy storage systems, and consumer electronics, lithium batteries constitute a strategically critical material for the global low-carbon transition of energy systems [1]. In recent years, technological innovation in the lithium battery sector has been highly active, with patent applications continuing to grow at an explosive rate, and China has risen to become the world’s largest source of lithium battery patents [2]. Patents serve as legal instruments for protecting technological innovation. They also act as core carriers for the diffusion of industrial technology and the circulation of innovation resources. Nevertheless, a large proportion of lithium battery-related patents remain dormant for extended periods, and the low commercialization rate limits their conversion into industrial competitiveness [3]. Information asymmetry between supply and demand represents the central obstacle underlying this challenge: enterprises struggle to accurately identify target patents that meet their specific technical needs from patent databases that are vast in scale and complex in structure, while patent holders find it equally difficult to proactively reach potential assignees with genuine technical demands [4]. Against this backdrop, intelligent patent recommendation systems oriented toward enterprise technology demands have emerged as an important technical pathway for promoting patent technology transfer and commercialization and for unlocking the value of patent resources [5].

With the widespread use of patent data in technology intelligence research, scholars have examined patents in the fields of new energy and energy storage from multiple perspectives. Some studies have applied patent bibliometrics, topic identification, and patent network construction to reveal technological trends and innovation structures. Wang et al. [6] constructed patent collaboration, knowledge, and transfer networks in the Chinese energy storage industry. They found that the structure of the knowledge network significantly influenced patent transfer activities. Jian et al. reported that a firm’s position in the patent collaboration network and its innovation environment affected patent transfer behavior [7]. In addition, some studies adopted the perspective of technological innovation systems and examined the interactions among publications, patents, and technical standards in knowledge diffusion. These studies showed that such knowledge carriers jointly form key channels of knowledge flow within innovation systems [8]. At the level of knowledge contribution in lithium battery innovation, Feng et al. [9] systematically analyzed the transformation pathways from scientific knowledge to patent technologies from an industrial chain perspective. They found that knowledge flows were most intensive in electrode materials, electrolytes, and battery management systems. In terms of methodological approaches, Du et al. [10] proposed an ecological efficiency evaluation method that integrates multiple models. Their results showed that green technological innovation significantly improves urban ecological efficiency. Overall, these studies highlight the important role of patent data in understanding technological innovation from multiple perspectives.

Despite the continuous growth in patent output in the new energy sector, many patents have not been effectively commercialized. As a result, patent transfer and technology commercialization have become important research topics. Existing studies have mainly focused on the structure of technology transfer networks, patent value evaluation, and transfer prediction. Li et al. [11] constructed knowledge transfer networks in patent-intensive industries and analyzed their evolution and determinants. They found that regional innovation capability and demand structure significantly influenced knowledge transfer networks. Cai et al. [12] examined patent transfer from a regional innovation perspective. By modeling transfer relationships among firms, research institutions, and individuals, they revealed the structural characteristics of technology transfer and evaluated its impact on regional innovation performance. Moreover, patent transfer has been regarded as a key linkage mechanism in regional technology transfer networks. Its scale and structure have a significant impact on the efficiency of innovation networks [13]. At the methodological level, Liu et al. [14] applied machine learning and transfer learning methods to predict patent transfer behavior. Their approach provided support for technology commercialization decisions. In addition, the technological and economic value of green technology patents has been shown to significantly affect their likelihood of transfer [15]. These studies deepen the understanding of patent commercialization mechanisms at a macro level. However, systematic approaches for accurately matching enterprise-specific technological demands with patent supply remain limited.

To address the information asymmetry between enterprise technological demands and patent supply, researchers have increasingly explored patent supply–demand matching and recommendation methods. Early studies relied on keyword matching and text similarity approaches. Methods such as TF-IDF were used to measure similarity between patent texts and support retrieval and recommendation. With the development of natural language processing, researchers introduced semantic embedding models such as Word2Vec and Doc2Vec. These methods later evolved into contextual representation approaches based on BERT, which improved the accuracy of semantic matching [16,17,18,19,20]. Beyond text-based methods, some studies constructed heterogeneous information networks to capture relationships among patents and technological actors. For example, Xu et al. [21] proposed a recommendation method based on heterogeneous information networks. They used meta-path counting to achieve cross-domain recommendation between scientific publications and patents. At the same time, deep learning methods have been applied to patent recommendation and technology prediction tasks. Models such as recurrent neural networks and convolutional neural networks have been used to extract deep semantic features from patent texts [22,23,24,25].

In recent years, the development of knowledge graphs and large language models has led to growing interest in patent recommendation based on structured knowledge representation. By constructing patent knowledge graphs, researchers can represent technical entities, relationships, and innovation actors in a structured form. This supports semantic retrieval and relational reasoning over complex technical knowledge. For example, Yang et al. [25] proposed a patent knowledge recommendation framework that integrates knowledge graphs with large language models. Their approach combined fine-grained technical ontologies with graph representation learning to enable semantic mining and recommendation. Retrieval-Augmented Generation (RAG) provides a new paradigm for knowledge-intensive information retrieval. It integrates external knowledge retrieval with the generative capability of large language models. This enables reasoning over multi-source knowledge and improves retrieval accuracy in complex query scenarios [26]. The integration of knowledge graphs and RAG has shown strong performance in industrial knowledge reasoning tasks. Bahr et al. embedded knowledge graphs into a RAG framework and improved semantic question-answering in industrial failure analysis [27]. Wei et al. [28] proposed a query-aware multi-path knowledge graph fusion method. By constructing multi-hop subgraphs, their approach enhanced retrieval relevance and improved the generation quality of large language models. These advances suggest that combining knowledge graphs and RAG has strong methodological potential for patent recommendation. Applying these techniques to enterprise-oriented technological demand scenarios can provide new directions for intelligent supply–demand matching.

However, when existing methods are applied to patent recommendation tasks driven by real enterprise technological demands, three systematic limitations remain. First, existing research lacks a differentiated processing framework that accounts for variation in demand expression completeness. The vast majority of methods treat enterprise technology demands as homogeneous text inputs, failing to distinguish the fundamental structural differences between complete demands with full element coverage and incomplete demands with missing information. As a result, a single recommendation strategy cannot adequately accommodate the diversity of enterprise demand scenarios [29]. Second, the semantic gap between supply and demand has not been effectively resolved. Enterprise demands are predominantly expressed in colloquial and non-standardized language, while patent texts adopt highly formalized technical language. Existing methods either rely on surface-level lexical matching or introduce complex ontological structures, and none of them construct a structured matching pathway that bridges the semantic gap by grounding the approach in the correspondence between demand-side elements and patent supply-side elements [30]. Third, the interpretability of recommendation results is generally insufficient. Most deep learning methods target end-to-end prediction and cannot provide explicit element-level correspondences, making it difficult for enterprises to understand the specific relationship between a recommended patent and their own demand, thereby limiting the practical utility of recommendation systems in real-world transfer decisions [31]. These limitations are particularly salient in the lithium battery domain, which is characterized by high technological intensity and highly refined supply chain specialization, and there is an urgent need for patent recommendation methods tailored to differentiated demand characteristics.

Taking the lithium battery technology domain as the research context, this study proposes a knowledge graph-based differentiated patent recommendation framework driven by enterprise technology demands. The framework first constructs a content element system covering five element types (material, method, efficacy, product, and application) on both the supply and demand sides. Enterprise technology demands are then classified into complete demands and incomplete demands according to the degree of element coverage. For each demand, a corresponding patent supply knowledge graph is constructed using patents potentially matching the enterprise’s requirements, enabling structured organization and semantic relational representation of patent supply information. On this basis, differentiated patent recommendation methods are designed for each demand type.

For complete demands, a Precision Recommendation Method for Complete Technological Demands is proposed. This method integrates BERT semantic encoding with TransE structural modeling and introduces an RAG-based intelligent constraint retrieval mechanism. Two conditions, namely the same patent node constraint and priority for full five-element coverage, are embedded into the retrieval instructions to prevent pseudo-matching caused by cross-patent element conflation. The final recommendation of a single optimal patent is achieved through a three-dimensional weighted scoring scheme incorporating supply–demand content conformity, patent element structural coherence, and element coverage.

For incomplete demands, a Fuzzy Recommendation Method for Incomplete Technological Demands is proposed. This method overcomes the information bottleneck of relying solely on the demand text itself by incorporating three types of external data, namely enterprise industry background, historical patent data, and technical job recruitment information, to supplement the demand category boundaries. Building on this, the sparse demand element information is expanded into semantically richer retrieval conditions through demand category keyword composite fitness ranking and augmented query context construction. Combined with knowledge graph-constrained retrieval and large language model-based intelligent filtering, a diverse set of candidate patents is generated through dual-dimensional scoring based on supply–demand content conformity and demand category fitness, providing recommendation results for ambiguous demands that balance broad coverage with directional accuracy.

At the empirical level, enterprise technology demand texts are collected from 12 provincial science and technology achievement transfer and transformation platforms. Following keyword-based screening and manual verification, 197 valid demands in the lithium battery domain are identified. Using the demand publication date as the reference point, patent assignment records of demand-submitting enterprises occurring after the publication of their demands are retrieved. Through a combination of BERT semantic similarity computation and manual assessment, 25 patent assignment cases are ultimately confirmed as reasonably attributable to technology demand-driven motivations, forming the empirical validation dataset. By comparing recommendation results against actual transfer behavior, the effectiveness of both recommendation methods is evaluated, providing systematic methodological support for enterprises in the lithium battery domain to efficiently identify and acquire external technological resources and to promote patent technology transfer and commercialization.

2. Enterprise Demand Classification and Patent Supply Information Organization

2.1. Enterprise Demand Classification Based on Content Elements

2.1.1. Identification and Construction of Content Elements for Enterprise Technology Demands

Structural-functional theory holds that elements are the fundamental units constituting objective entities and serve as necessary conditions for their existence and operation. Different elements fulfill specific functions and form a complete system through relatively stable interconnections [32]. Although enterprise technology demands vary considerably in their expression, they can be decomposed from an element-based analytical perspective into a set of basic units that share inherent logical relationships.

Existing research has primarily used patents and academic publications as analytical texts, from which a general element classification framework has been derived covering dimensions such as product, method, material, efficacy/performance, technical attributes, and application domain [33]. However, such frameworks are centered on revealing the principles and implementation pathways of technological innovation, and do not align well with how enterprises actually articulate their technology demands. When expressing technical needs, enterprises typically focus on specific bottlenecks encountered in production, improvement targets, and functional goals, rather than systematically describing innovation principles or technical implementation mechanisms. Consequently, directly applying element frameworks developed from scientific literature cannot accurately capture the content characteristics of enterprise technology demands. It is therefore necessary to develop a content element classification system that better reflects how enterprise demands are expressed in real-world texts.

To construct an element classification system that better reflects the expression characteristics of enterprise technological demands, this study conducted a systematic content analysis based on authentic enterprise demand data. Enterprise technological demand texts were collected from public service platforms for technology transfer and commercialization established in twelve Chinese provinces, including Anhui, Fujian, Gansu, Hebei, Henan, Heilongjiang, Hunan, Jilin, Jiangxi, Shanxi, Shaanxi, and Yunnan. A total of 5000 enterprise technological demand records in the new energy sector were obtained. The new energy sector was selected as the analytical sample because it encompasses multiple technological directions, including lithium batteries, photovoltaics, energy storage, and related materials. The diversity of technological fields in this sector results in a wide range of demand types, which can more comprehensively reflect the common technological problems faced by enterprises during technological innovation and the ways in which these demands are expressed. Constructing the enterprise technological demand element system based on new energy demand data helps identify the general structural characteristics of enterprise technological demands and provides a structured analytical framework for subsequent research on technological demand identification and patent supply–demand matching in the lithium battery field.

The analytical process of this study consisted of two stages. In the first stage, the BERTopic model was applied to conduct topic modeling and semantic clustering on enterprise technological demand texts in order to identify recurring technological themes and representative semantic patterns. BERTopic integrates Transformer-based semantic embedding models with the class-based TF–IDF (c-TF-IDF) weighting algorithm. This integration preserves high-value semantic information during topic extraction and improves topic distinctiveness and interpretability. In this study, the text embedding model paraphrase-multilingual-MiniLM-L12-v2 was used to generate semantic vector representations of the demand texts. The clustering parameters were set as n_neighbors = 15, n_components = 2, min_cluster_size = 80, and min_samples = 10. These parameter settings achieve a balance between topic coherence and clustering granularity and enable the model to effectively capture representative semantic patterns in enterprise technological demands. The clustering results are shown in

Table 1. Thematic clustering results of enterprise technology demands.

Topic	Representative Keywords	Topic	Representative Keywords
1	reduction furnace’, ‘magnesium metal’, ‘neodymium–iron–boron’, ‘oil shale’, ‘product quality’	8	‘high-temperature resistance’, ‘thermal stability’, ‘anti-aging’, ‘flame retardancy’, ‘glass wool’
2	‘copper alloy’, ‘rare earth metals’, ‘cemented carbide’, ‘non-ferrous metals’, ‘diamond’	9	‘aluminum alloy’, ‘aluminum profile’, ‘bauxite’, ‘alumina’, ‘gibbsite’
3	‘dimethyl sulfoxide ‘, ‘boron nitride’, ‘calcium carbonate’, ‘magnesium alloy’, ‘caprolactam’	10	‘lithium battery’, ‘lithium-ion’, ‘metallic lithium’, ‘power battery’, ‘electrolyte’
4	‘tensile strength’, ‘boron carbide’, ‘elongation’, ‘elongation at break’, ‘recovery rate’	11	‘concrete’, ‘accelerator’, ‘steel fiber’, ‘fly ash’, ‘admixture’
5	‘sewage treatment’, ‘wastewater treatment’, ‘biodegradation’, ‘water-reducing agent’, ‘harmless treatment’	12	‘polylactic acid (pla)’, ‘copolymer’, ‘acrylic acid’, ‘tetrahydrofuran’, ‘lactide’
6	‘waterproof coating’, ‘waterproof materials’, ‘protective layer’, ‘polyurethane’, ‘sealant’	13	‘solar energy’, ‘magnetofluid’, ‘electromagnet’, ‘bevel gear’, ‘generator’
7	‘sensor’, ‘robot’, ‘automation’, ‘subsystem’, ‘fully automatic’	14	‘glass fiber’, ‘carbon fiber’, ‘cellulose’, ‘ceramic fiber’, ‘fibrous materials’

. The representative keywords of each topic mainly involve material types, performance indicators, industrial processes, product forms, and application scenarios.

In the second stage, representative demand texts within each topic cluster were further examined and manually coded to identify the structural common elements of enterprise technological demands. Although enterprises may differ in their expression styles and technical details when describing technological demands, the core content of these demands exhibits relatively stable structural characteristics. Through the synthesis and structural analysis of the semantic content of demand texts, enterprise technological demands can be classified into five fundamental elements: materials, methods, efficacy, products, and applications. The specific meanings and definitions of these five elements are presented in

Table 2. Content element framework for enterprise technology demands.

Element Type	Definition	Representative Expressions
Material	Key raw materials, auxiliary materials, functional components, or technical processing objects on which the enterprise relies to achieve specific technical objectives.	“Development of a new composite material is needed”; “Replace the substrate to improve corrosion resistance”
Method	Technical implementation pathways or operational approaches, including process flows, preparation methods, operational steps, and technical realization procedures.	“Optimize the production process”; “Improve the detection method”; “Develop a preparation process”
Efficacy	Specific functional or performance objectives the enterprise expects to achieve through technological means, including performance enhancement, energy efficiency optimization, and quality improvement.	“Increase catalytic efficiency”; “Enhance thermal stability”; “Meet environmental emission standards”
Product	The form of technological output the enterprise seeks to obtain, encompassing both physical products such as equipment, devices, and components, and virtual products such as software systems and algorithmic models.	“Develop a new cell structure”; “Develop a battery management software system”
Application	The specific deployment scenarios or applicable industry contexts of the technology.	“Applicable to the new energy vehicle sector”; “Targeted at industrial wastewater treatment scenarios”

.

Compared with technological element systems derived from scientific literature, the five-element framework proposed in this study better reflects the practical concerns of enterprises when expressing technological demands. Enterprises usually describe their demands in terms of technological bottlenecks encountered in production practice, directions for improvement, and intended application goals. The five elements proposed in this study are not derived from industry-specific terminology. Instead, they represent an abstract generalization of the structural components of technological problem descriptions and capture the general composition of technological problems. Therefore, the five-element framework provides a universal analytical framework for the structured representation and analysis of enterprise technological demands. It also establishes an important structured semantic foundation for subsequent research on demand identification, supply–demand matching, and patent recommendation.

2.1.2. Classification of Enterprise Technology Demands Based on Element Completeness

In an ideal scenario, a fully expressed enterprise technology demand should encompass all five content element types, thereby providing a complete representation of the enterprise’s technical requirements. In practice, however, the extent to which demand information is complete is influenced by the enterprise’s level of technical understanding, its willingness to disclose information, and its confidentiality strategies, resulting in significant variation in element coverage across demand texts [34].

In the lithium battery domain, for example, large enterprises with substantial technological expertise can typically articulate the required materials, methods, efficacy targets, products, and applications in a well-defined manner, yielding fully specified demands. By contrast, small and medium-sized enterprises that are still in the exploratory phase of technology development, or that seek to protect sensitive business information, often express their demands only in terms of general improvement directions or functional goals. A representative example would be a demand stating that the enterprise hopes to extend battery cycle life and resolve the problem of capacity degradation, with no further element-level detail provided. Such demands are noticeably incomplete in terms of element coverage.

Based on the above analysis, this study classifies enterprise technology demands into two types according to the degree of element coverage.

The first type is complete technology demands, which contain all five content element types: material, method, efficacy, product, and application. Such demands articulate the technical problem with clarity and specificity, often including quantifiable performance parameters or technical indicators. This indicates that the enterprise has developed a systematic understanding of its technical needs and possesses strong demand articulation capabilities. Complete technology demands provide sufficient element-level constraints for patent recommendation and are well-suited for precise patent-to-demand matching.

The second type is incomplete technology demands, which lack one or more of the five element types and contain only one to four elements. Such demands are typically vague in their description of the technical problem, lack element-level detail, and do not include quantifiable performance indicators, making precise recommendation through direct element correspondence unfeasible. Incomplete demands are highly prevalent in practice, particularly in rapidly evolving fields such as new energy, and represent one of the primary challenges currently facing patent recommendation research.

The fundamental difference in information completeness between these two demand types determines that patent recommendation strategies cannot be uniformly applied. Instead, differentiated matching methods must be designed for each demand type. This recognition constitutes the core logical premise underlying the patent recommendation approach proposed in this study.

2.1.3. Identification Methods for Enterprise Technology Demands

Enterprise technology demand texts are unstructured natural language in which element information is embedded in a colloquial and non-standardized manner across different sentence segments. Rule-based matching methods are ill-suited to handle the complexity and variability of such expressions. This study therefore designs distinct element identification strategies tailored to each demand type.

For complete technology demands, the element identification task is formulated as a named entity recognition (NER) problem. ChatGPT-5 is introduced as the identification tool, and prompt engineering is employed to automatically extract entities corresponding to the five element types [35]. A standardized prompt template is designed incorporating four core components: role specification, task objective, output format constraints, and entity boundary definitions (see Appendix B, Figure A1 for the prompt template). Demand texts are combined with the prompt template and submitted to the model item by item to obtain the element content for each entry. This approach requires no large-scale domain-annotated training data and demonstrates strong adaptability to non-standard terminology and colloquial expressions. It accurately extracts entity information across all five element types and provides structured demand-side input for subsequent patent recommendation methods [36]. In this process, ChatGPT-5 was used to generate entity extraction outputs based on the designed prompt templates. All extracted results were manually reviewed and corrected by the authors to ensure accuracy and reliability.

For incomplete technology demands, the absence of one or more element type means that the demand text alone is insufficient to support effective patent recommendation. To address this, the study proposes a multi-source data-driven approach to supplement the categorical scope of incomplete demands. Rather than attempting to reconstruct each missing element individually, this method introduces three types of external data that are closely related to enterprise technological activities to define the boundaries of technology demand categories and constrain the range of potentially relevant technologies at a directional level. The three data sources are as follows. The first is enterprise industry background [37], comprising industry classification and primary business scope, which is used to identify the technology direction category and reflects the industrial track in which the enterprise operates. The second is enterprise patent data [38], consisting of historically filed patents from which high-weight technical keywords are extracted using TF-IDF to define the technology base category and reflect the enterprise’s accumulated technological competencies. The third is enterprise technical recruitment information [39], referring to the skill requirements listed in technical job postings, which is used to identify capability demand categories and reveals the technology directions the enterprise currently seeks to acquire. These three data sources are encoded using BERT semantic vectors, and cosine similarity is computed to construct a demand category matrix. This matrix is combined with the already identified elements to form the input representation for the incomplete demand-oriented patent recommendation method.

2.2. Organization of Patent Supply Information Based on Knowledge Graphs

On the supply side, patent texts carry rich technical content and serve as an important source of technological solutions for enterprises. To ensure that supply and demand can be matched at the element level, this study first performs element identification and analysis on patent supply content. A dataset of 100,000 patents in the lithium battery technology domain was compiled from the IncoPat patent database. The dataset covers multiple technical directions including cathode and anode materials, electrolyte systems, separator technologies, preparation processes, cell structural design, and typical application scenarios, offering strong technical representativeness and content diversity that collectively reflect the content distribution characteristics of patent supply in the lithium battery domain.

BERTopic, an unsupervised semantic clustering model, is applied to identify content elements from the patent corpus [40]. Through topic categorization and content integration, patent supply content can be organized into five element types as follows.

The first is the material element, which refers to the physical substance entities underlying the patented technology, such as active cathode and anode materials, conductive agents, electrolytes, separators, and surface modification agents. The second is the method element, which refers to the operational procedures, process techniques, or implementation pathways designed in the patent to achieve specific technical objectives, such as preparation, assembly, heat treatment, coating, and sintering processes. The third is the efficacy element, which refers to the quantifiable performance outcomes achieved by the patented technical solution, such as improvements in energy density, rate capability, thermal stability, and cycle life. The fourth is the product element, which refers to the functional carriers of the patented technology that embody its technical value in physical or digital form, such as BMSs, cell modules, and battery pack architectures. The fifth is the application element, which refers to the specific deployment scenarios, operating environments, or service domains of the patented technology, such as electric vehicles, portable devices, energy storage systems, and aerospace equipment.

These five patent supply element types correspond one-to-one with the five content element types on the enterprise demand side, indicating that the element framework through which enterprises articulate technical problems and the element framework through which patents present technical solutions share an inherent structural consistency. This structural correspondence provides a theoretical basis for establishing supply–demand relationships and designing patent recommendation methods.

Having established the patent supply element framework, the question of how to efficiently organize large-scale patent supply information in a structured manner directly determines the matching efficiency and precision of the recommendation method. Conventional keyword retrieval and text indexing approaches face three inherent limitations when applied to patent supply–demand matching. First, these methods rely on surface-level lexical overlap and are unable to capture cross-semantic associations between patent language and enterprise demand expressions, making the vocabulary gap a particularly salient problem. Second, patent information spans multiple element dimensions covering material, method, efficacy, product, and application, and flat text indexing cannot effectively represent the structural relationships among different element entities, thus failing to support multi-dimensional element-level matching. Third, for incomplete demands, missing element information severely impairs the recall of keyword-based retrieval, as the available elements alone are insufficient to locate potentially relevant patents [41].

Knowledge graphs offer effective solutions to these challenges through their distinctive strengths in semantic modeling, structured representation, and relational reasoning. By representing information as entity–relation–attribute triples, knowledge graphs transform unstructured patent texts into structured semantic networks. Their advantages manifest in three respects.

The first is semantic alignment capability. By extracting patent supply elements as standardized entity nodes and connecting patents to various element types through semantic relations, knowledge graphs establish a structured correspondence between patent supply and enterprise technology demands at the element level, thereby enabling semantic-level matching that transcends lexical differences [42].

The second is the structured integration of multi-dimensional elements. A knowledge graph can organize the five element types of any given patent within a patent node–relation edge–element node network structure. This design ensures that each element is independently retrievable while maintaining the integrity constraint that all elements belong to the same patent, effectively preventing false matches caused by cross-patent element conflation.

The third is semantic connectivity and reasoning capability. The relational network within a knowledge graph supports not only direct entity retrieval starting from demand elements, but also multi-hop reasoning across entities via relational paths. This provides a structured reasoning channel for inferring relevant patents from partially known elements in incomplete demand scenarios [43].

Guided by the principle of supply–demand element mapping, this study designs the entity and relation schema of the knowledge graph with the patent number as the core hub node, along with five element entity types: material, method, efficacy, product, and application. At the relation level, semantic relations are defined between patents and each element entity type, as shown in

Table 3. Entity and relation schema of patent supply knowledge graph.

Type	Name	Design Rationale
Entity	Material	Corresponds to the “material” element of enterprise demands
	Method	Corresponds to the “method” element of enterprise demands
	Efficacy	Corresponds to the “efficacy” element of enterprise demands
	Product	Corresponds to the “product” element of enterprise demands
	Application	Corresponds to the “application” element of enterprise demands
Relation	Patent_Contain_Material	Association between patents and the “material” element
	Patent_Use_Method	Association between patents and the “method” element
	Patent_Achieve_Efficacy	Association between patents and the “efficacy” element
	Patent_Produce_Product	Association between patents and the “product” element
	Patent_Apply_Application	Association between patents and the “application” element

. Regarding data construction, differentiated entity extraction strategies are adopted according to the semantic characteristics and extraction difficulty of each entity type. Efficacy entities are extracted directly from the standardized technical efficacy phrase field in the IncoPat database. Material, method, product, and application entities are extracted from patent abstracts using a three-stage strategy comprising initial extraction, manual verification, and few-shot augmentation, with ChatGPT-5 as the core tool and few-shot prompting employed to improve recognition accuracy for domain-specific terminology. Specifically, ChatGPT-5 was used to perform the initial entity extraction and few-shot augmented recognition stages. The manually verified outputs in the intermediate stage were conducted by the authors, who also reviewed all final extraction results and take full responsibility for the accuracy of the structured data used in the knowledge graph. Relations are constructed using a rule-driven method that automatically generates relation triples based on the co-occurrence pattern of patent number, entity, and relation type. After deduplication and consistency verification, the triples are imported into the Neo4j graph database for storage and visualization. The overall ontology structure of the patent supply knowledge graph is illustrated in Figure 1.

3. Patent Recommendation Methods for Different Types of Enterprise Technology Demands

3.1. Classification of Patent Recommendation Methods Based on Demand Types

The design of patent recommendation algorithms is grounded in supply–demand matching, with its core being the establishment of supply–demand content associations and patent positioning through the patent supply knowledge graph. Recommendation algorithms encompass not only patent retrieval and screening but also the modeling and reasoning process of supply–demand correspondence relationships. Building upon the demand classification framework established in Section 2, which categorizes enterprise technological demands into complete demands and incomplete demands, this section proposes differentiated patent recommendation methods tailored to the distinct characteristics of each demand type. These two demand types exhibit significant differences in element types and expression characteristics. The diversity of enterprise demands determines that recommendation algorithms cannot adopt a single approach but must be designed according to differentiated supply–demand characteristics.

Combining demand types with supply–demand association relationships, this study proposes that patent recommendation algorithms must satisfy differentiated matching requirements in their design and accordingly constructs corresponding recommendation modes:

1. Precise Matching for Complete Demands

Complete demands express all five content element types—Material, Method, Efficacy, Product, and Application. Each element explicitly describes demand content and constraint conditions. The core requirement is to identify optimal patent supply that highly aligns with demand elements and can directly address technical problems. Patent recommendation algorithms for such demands belong to precise matching. Through multiple matching condition constraints including element coverage, semantic correspondence, and metric alignment, they achieve optimal matching between recommendation results and enterprise demands. This recommendation type emphasizes high precision and high certainty of patent recommendation results, requiring algorithms to provide explicit responses to each demand element type proposed by enterprises. This represents the most direct and deterministic matching relationship in supply–demand associations.

2. Fuzzy Matching for Incomplete Demands

Incomplete demands are characterized by ambiguous expression and element deficiency, containing only partial element information and making it difficult for a single patent to accurately respond to enterprise demands. Patent recommendation algorithms for such demands belong to fuzzy matching. By integrating multi-source data to supplement demand categories and combining demand problem identification with supply solution reasoning, these algorithms transform demands into retrievable patent supply categories. The core objective of this recommendation algorithm is not to find the most precise patent supply solution but rather to form diversified candidate patent sets based on demand categories. By expanding the supply–demand matching space through demand categories, these algorithms provide demand parties with multiple patent supply solutions for comparative selection, achieving a matching relationship where one demand corresponds to multiple patent recommendation solutions.

3.2. Precision Recommendation Algorithm for Complete Demands

This study defines the recommendation algorithm for identifying patent supply solutions for complete demands as the “Precision Recommendation Algorithm for Complete Technological Demands” (PR-CTD). PR-CTD aims to identify the most precise patent supply solutions from the patent supply knowledge graph for individual complete enterprise demands. During the recommendation process, it must ensure that recommended patents achieve semantic correspondence and quantitative coverage across all five element types—Material, Method, Efficacy, Product, and Application. Under the objective of precise patent recommendation, PR-CTD emphasizes three aspects: (1) Element matching results must consist of five element entities from the same patent, avoiding “pseudo-matching results” formed by cross-patent element splicing; (2) Dual constraints of supply–demand element semantic matching and patent element type completeness must be simultaneously satisfied, ensuring recommendation results are reasonable in both element content and element quantity; (3) Patent recommendation results must possess interpretability, providing rationales for patent recommendations and offering matching bases for both supply and demand parties.

Building upon the traditional knowledge graph embedding model TransE and integrating LLMs with enterprise demand lexical set, this study employs Retrieval-Augmented Generation (RAG) technology, enabling LLMs to transform demand texts into retrieval statements under demand constraints, compensating for deficiencies of traditional Trans models [44]. Compared with traditional single Trans models, this method offers the following advantages:

(1) Enhanced semantic understanding capability. Addressing TransE’s limitation of primarily relying on graph structure learning with limited semantic understanding, this study proposes a dual-vector encoding method combining BERT and TransE. BERT leverages semantic knowledge acquired during pre-training to effectively identify synonyms, hierarchical terms, and functionally equivalent expressions in supply–demand texts, enhancing the model’s understanding of natural language semantic differences. TransE retains its advantage in modeling knowledge graph structural relationships [45]. The fusion of both achieves unification of semantic understanding and structural modeling, resolving the issue where traditional TransE methods identify synonymous expressions as different entities, thereby affecting matching accuracy.

(2) Strengthened global consistency constraint mechanism. Addressing the lack of global constraints in Trans-series models, this study designs an intelligent constraint retrieval mechanism based on RAG. During the query generation stage, “same patent node constraints” are embedded in the LLM’s retrieval instructions. During the candidate patent generation stage, retrieval conditions for complete coverage of all five element types are prioritized. The enterprise demand lexical tree provides semantic support for RAG, ensuring constraint conditions are accurately transmitted during semantic expansion. This method simultaneously considers both element content matching and element completeness constraints, ensuring recommended patents not only accurately correspond to enterprise demand elements at the semantic level but also completely cover all five element types in content. The enterprise demand lexical tree provides semantic support for RAG, ensuring that constraint conditions are accurately transmitted during semantic expansion. Specifically, the demand lexical tree is constructed using the five demand elements as the first-level nodes, forming a hierarchical demand terminology structure. Core demand terms extracted from enterprise technological demand texts are first organized through manual annotation, and then expanded using a large language model to generate semantically related expressions such as synonyms, alternative terms, and hypernyms. Based on this structure, each identified demand element $e_i$ can be expanded into an extended semantic set $E_{ext}\left(e_i\right)$, which is used during the RAG retrieval process to improve semantic coverage between demand expressions and patent texts. This method simultaneously considers both element content matching and element completeness constraints, ensuring recommended patents not only accurately correspond to enterprise demand elements at the semantic level but also completely cover all five element types in content.

(3) Effective handling of data sparsity and sampling bias issues. Traditional deep learning models typically rely on large-scale training when addressing data sparsity and sampling bias issues, resulting in low efficiency and high costs. This study introduces RAG, transforming the training-dependent mode into an intelligent reasoning mode. Within this framework, the enterprise demand lexical tree serves as a terminology set for demand expansion, providing semantic context for LLMs; LLMs function as intelligent reasoning engines, understanding deep semantics of demands and generating high-quality retrieval statements; and the knowledge graph serves as the patent supply knowledge base, providing relational constraints for supply–demand matching. Through this triple assurance mechanism, RAG can transform demands into precise patent queries, effectively addressing sampling bias and sparse data issues in traditional methods.

The PR-CTD scheme is illustrated in Figure 2 and comprises five main stages:

Stage 1: Demand Parsing and Element Expansion

Demand parsing is the initial step of the PR-CTD algorithm, with the objective of identifying five element types from complete enterprise demands. The complete enterprise demand text is denoted as D, and five element types are extracted, Material, Method (Approach), Efficacy, Product, and Application (Use), as shown in Formula (1):

$$E_D=\{M_D,A_D,E_D,P_D,U_D\}$$

(1)

where $M_D$ represents the material element set, $A_D$ represents the method element set, $E_D$ represents the efficacy element set, $P_D$ represents the product element set, and $U_D$ represents the application element set.

To address the semantic diversity issue in demand expression, the PR-CTD algorithm introduces the demand lexical tree $T$ for element expansion. For each original element $e_i$, based on the hierarchical structure and semantic associations of the demand lexical tree $T$, an expanded element set $E_{ext}\left(e_i\right)$ is generated.

Stage 2: Dual-Vector Encoding and Fusion

To fully utilize element semantic information and structural information from the patent supply knowledge graph, the PR-CTD algorithm employs a dual-vector encoding strategy combining BERT and TransE, comprehensively integrating element semantic and structural information. BERT performs deep semantic encoding of expanded elements through a pre-trained language model, obtaining vector representations containing contextual information. TransE associates elements with corresponding entities in the knowledge graph, obtaining embedding vectors that reflect structural relationships among elements. Finally, these two vector types are unified through a weighted fusion strategy. For each element $e_i$, its BERT semantic vector and TransE structural vector are obtained respectively.

BERT Semantic Vector Encoding: Element text is input into the pre-trained BERT model, utilizing the hidden state corresponding to the [CLS] token in the final layer as the semantic representation of the element (Formula (2)):

$$v_{hert}(e_i)=\mathrm{BERT}{(e_i)}_{[CIS]}$$

(2)

TransE Structural Vector Acquisition: Element $e_i$ is mapped to its corresponding entity node ${\tilde{e}}_i$ in the patent supply knowledge graph, and its structural embedding vector pre-trained in the TransE model is extracted (Formula (3)):

$$v_{trans}(e_i)=\mathrm{Trans}\mathbb{E}({\tilde{e}}_i)$$

(3)

To achieve effective fusion of BERT semantic vectors and TransE structural vectors, it is necessary to address dimension alignment and weight allocation issues. The semantic vectors output by the BERT model have a dimension of 768, while the TransE model embedding vectors have a dimension of 128. The two cannot directly undergo numerical operations. Therefore, a linear transformation matrix $\mathbf{W}\in {\mathbb{R}}^{768\times 128}$ and fusion weight $\alpha$ are introduced, employing normalized weighted fusion as shown in Formula (4):

$$v_{fised}(e_i)=\alpha \cdot v_{bert}(e_i)+(1-\alpha )\cdot W\cdot {\mathit{v}}_{trans}(e_i)$$

(4)

where ${\mathbf{v}}_{fused}\left(e_i\right)$ represents the final fused vector representation of element $e_i$. Both the fusion weight $\alpha$ and transformation matrix $\mathbf{W}$ are set as trainable parameters to construct a joint optimization objective based on the candidate patent set, as shown in Formula (5):

$$\{{\alpha }^{\ast },W^{\ast }\}=\arg \underset{\alpha ,W}{\max }\sum _{p_i\in C}Score(p_i,\alpha ,W)$$

(5)

Iterative updates are performed through the Adam optimizer with a learning rate set to 0.01. The objective of each iteration is to maximize the comprehensive matching score of the candidate patent set, synchronously updating the fusion weight parameter $\alpha$ and matrix $\mathbf{W}$ parameter through backpropagation. The comprehensive score of candidate patents is defined as the sum of weighted scores across three dimensions for all candidate patents: supply–demand element conformity, patent element structural rationality, and patent element coverage, used to measure the overall matching quality between the entire candidate set and enterprise technological demands. To ensure the meaningfulness of the fusion weight, $\alpha$ is constrained to the range [0, 1] after each parameter update.

Figure 3 illustrates the convergence trajectory of parameter $\alpha$ during the iteration process and corresponding performance changes. Experimental results indicate that the $\alpha$ value continuously decreases over 20 iterations and converges to 0.367, with the corresponding comprehensive matching score improving from 13.80 to 14.91. After the 12th iteration, the rate of $\alpha$ value decrease slows significantly, indicating that the model gradually approaches an optimal state. The improvement magnitude of the comprehensive matching score becomes more gradual, with diminishing marginal returns. Considering the limited marginal benefit of further training, this study establishes $\alpha =0.367$ at the 20th iteration as the optimal weight.

The optimal weight ratio determined through experimentation shows a BERT semantic information contribution of 36.7% and a TransE structural information contribution of 63.3%. This result indicates that in patent recommendation tasks, the contribution of structural information exceeds that of pure semantic information, aligning with the characteristics of patents having high structural organization and relatively explicit entity relationships.

Stage 3: Intelligent Constraint Retrieval Based on RAG

After obtaining demand vector representations, the PR-CTD algorithm employs RAG to screen candidate patent supply solutions from the patent supply knowledge graph. The objective of this stage is to efficiently screen a small-scale candidate patent set that highly matches enterprise technological demands from a large-scale patent collection, ensuring both retrieval precision and enhanced computational efficiency for subsequent stages. Unlike traditional query approaches, RAG integrates graph retrieval with LLM reasoning through a three-stage process of “retrieval–augmentation–generation,” achieving transformation from semantic retrieval to intelligent optimization. The prompt instructions for RAG-based candidate patent screening are provided in Appendix B, Figure A2 (Prompt template for candidate patent screening).

(1) Retrieval Stage

The objective of this stage is to execute constraint retrieval based on fused vectors in the patent supply knowledge graph, obtaining an initial candidate patent set with complete structure and relevant content.

Based on the fused vectors ${\mathbf{v}}_{fused}\left(e_i\right)$ obtained in Stage 2, dual constraint retrieval is executed in the patent supply knowledge graph. CYPHER query statements are constructed in the Neo4j graph database to identify patents possessing all five elements through pattern matching and calculate similarity between the five-element fused vectors and patent element embedding vectors. Empirical results confirm that $\tau =0.7$ can identify element correspondence relationships while ensuring semantic accuracy; therefore, the similarity threshold ${\tau }_r=0.7$ is set as the patent retrieval threshold.

Constraint 1: CYPHER query statements are constructed to ensure that the five element types of candidate patents must originate from the same patent node, expressed as Formula (6):

$$\forall e_j\in \{m,a,e,p,u\},\exists !Paten{t}_k:(Paten{t}_k,r_j,e_j)\in \mathrm{KG}$$

(6)

$(Paten{t}_k,r_j,e_j)$ represents a triple in the patent supply knowledge graph $KG$, indicating that patent $Paten{t}_k$ is connected to element $e_j$ through relation $r_j$. Constraint 1 aims to avoid pseudo-matching issues arising from splicing elements across different patents, ensuring element completeness of recommendation results.

Constraint 2: Patent ranking based on vector similarity. Priority is given to retrieving patents covering all five element types. When this condition cannot be satisfied, the requirement is relaxed to covering at least four element types with mandatory inclusion of Method and Efficacy, expressed as Formula (7):

$$R=\{p_i\mid \mathrm{Coverage}(p_i)\ge 4\wedge \{\mathrm{method},\mathrm{efficacy}\}\subseteq \mathrm{Elements}(p_i)\}$$

(7)

$R$ represents the initial candidate patent set obtained from retrieval, providing input data for subsequent augmentation and generation stages. $Coverage(p_i)$ denotes the number of element categories covered by patent $p_i$, and $Elements(p_i)$ represents the set of element types contained in patent $p_i$.

(2) Augmentation Stage

The objective of the augmentation stage is to combine retrieved patents with demand information to construct a contextual environment for LLM reasoning. This enables the model to possess an information foundation for executing intelligent reasoning and provides knowledge support for subsequent intelligent screening.

The top 20 candidate patents are selected from retrieval set $R$, their five element types are extracted, and these are combined with demand element information to form the augmented context $T$. Context $T$ comprises two core components, expressed as Formula (8):

$$T=\{{\mathcal{D}}_{req},{\mathcal{P}}_{cand}\}$$

(8)

where ${\mathcal{D}}_{req}$ represents the complete enterprise demand element set, including demand elements and elements expanded based on the demand lexical tree, and ${\mathcal{P}}_{cand}$ represents the candidate patent set, containing the five element types for each patent.

(3) Generation Stage

The objective of this stage is to invoke LLMs to execute intelligent reasoning on the augmented context, screening out the patent set that most closely aligns with enterprise demands. This study employs Alibaba Cloud Tongyi Qianwen Plus (Qwen-Plus) as the reasoning model and combines prompt engineering to design task-oriented templates that guide the model in executing two analysis tasks:

(1) Element matching analysis: Based on information from the demand lexical tree, identify semantic relationships between demands and patents including synonyms, hypernym–hyponym relationships, and equivalent terms, evaluating the matching quality of the five element types in candidate patents.

(2) Comprehensive screening decision: Comprehensively evaluate each candidate patent across three dimensions: supply–demand element conformity (prioritizing patents with explicit element correspondence relationships), patent element structural rationality (prioritizing patents with strong technical associations among elements), and patent element coverage completeness (prioritizing patents covering all elements). Based on this evaluation, 10 patents are selected from the candidates to form candidate patent set $C$.

The entire RAG process achieves a complete workflow from vector retrieval through context augmentation to LLM intelligent optimization. The final output, candidate patent set $C$, serves as input for multi-dimensional matching score calculation in Stage 4, providing a high-quality candidate patent set for comprehensive scoring and precise recommendation.

Stage 4: Multi-Dimensional Matching Score Calculation

After obtaining the RAG-optimized candidate patent set $C$, the PR-CTD algorithm assesses the matching degree of each candidate patent $p_j\in C$ across three dimensions: supply–demand content conformity, patent element structural rationality, and patent element coverage, establishing a quantitative evaluation framework:

(1) Supply–Demand Content Conformity (For Complete Demands)

This metric measures the degree of semantic matching between elements of enterprise technological demands and patent supply. The cosine similarity between demand element vectors and patent element vectors is calculated, as shown in Formula (9):

$$S_{semantic}(D,p_i)={\displaystyle \frac{1}{5}}\sum _{j=1}^5\underset{e\in E_j(p_i)}{\max }\cos ({\mathit{v}}_D^{(j)},{\mathit{v}}_{p_i}^{(e)})$$

(9)

where ${\mathbf{v}}_D^{\left(j\right)}$ represents the vector representation of the $j$-th element type in the demand, and ${\mathbf{v}}_{p_i}^{\left(e\right)}$ represents the vector representation of the corresponding element in the patent.

(2) Patent Element Structural Rationality

This metric is used to verify the logical relationships among elements within the patent, ensuring that recommendation results are not only semantically similar but also structurally rational. TransE is utilized to verify the structural rationality of elements within patents:

$$S_{structure}(D,p_i)={\displaystyle \frac{1}{\left|R\right|}}\sum _{(p,r,e)\in R}\sigma (-\parallel h_p+r-t_e{\parallel }_2^2)$$

(10)

where ${\mathbf{h}}_p$ represents the TransE vector of the patent node, $\mathbf{r}$ represents the relation vector, ${\mathbf{t}}_e$ represents the element entity vector, and $R$ denotes the set of relations between the patent and its elements. $\sigma$ is the sigmoid function, used to convert the TransE distance metric into a standardized similarity score. The TransE model is based on the translation assumption $\mathbf{h}+\mathbf{r}\approx \mathbf{t}$. When vectors of the patent node, relation, and element entity satisfy this assumption, the Euclidean distance $\parallel {\mathbf{h}}_p+\mathbf{r}-{\mathbf{t}}_e{\parallel }_2^2$ approaches zero, indicating that triples with smaller distances receive higher structural rationality scores.

The core significance of this metric lies in verifying the causal relationships and logical rationality between patents and their elements—specifically, whether each element is essential to implementing the technology. This ensures that recommended patents not only exhibit semantic similarity to demands at the textual level but also match demands in element composition, thereby avoiding misrecommendations arising from superficial textual similarity.

(3) Patent Element Coverage

This metric is used to evaluate the completeness of patent supply’s response to enterprise technological demand elements. It counts the number of elements in candidate patents whose semantic similarity with the five demand element types exceeds a threshold and calculates the coverage ratio, as shown in Formula (11):

$$S_{coverage}(D,p_i)={\displaystyle \frac{\left|\right\{j|\exists e\in E_j(p_i),\cos (v_D^{(j)},v_{p_i}^{(e)})>\tau \}|}{5}}$$

(11)

This metric employs a threshold judgment mechanism: for each demand element type $j$, the algorithm checks whether a corresponding element exists in the patent with similarity exceeding threshold $\tau$. If such an element exists, the patent is considered to cover the $j$-th element type of the demand. The final coverage score is the ratio of the number of covered element categories to the total number of element categories. A patent with a coverage score of 1 indicates it contains all five element types. This study sets the similarity threshold $\tau$ to 0.7.

Stage 5: Comprehensive Scoring and Optimal Selection

Based on the scores from the three dimensions above, the PR-CTD algorithm employs a weighted linear fusion strategy to calculate the comprehensive matching score for candidate patents, as shown in Formula (12):

$$Score(p_i)={\beta }_1{S}_{semantic}(D,p_i)+{\beta }_2{S}_{structure}(D,p_i)+{\beta }_3{S}_{coverage}(D,p_i)$$

(12)

where ${\beta }_1,{\beta }_2,$ and ${\beta }_3$ are weight parameters satisfying ${\sum }_{i=1}^3{\beta }_i=1$. The weights are determined according to the functional roles of different evaluation dimensions in the patent recommendation process.

Supply–demand element similarity $S_{semantic}$ directly measures the semantic correspondence between enterprise technological demands and patent contents and therefore reflects the core relevance of recommended patents. Because semantic consistency between demand descriptions and patent technical information is the primary criterion for determining whether a patent can potentially address a technological need, this dimension plays the most critical role in recommendation quality. Therefore, β₁ is set to 0.5.

Patent element structural rationality $S_{structure}$ evaluates the logical consistency among internal patent elements based on the TransE model. This metric verifies whether the relationships among materials, methods, and technical effects within a patent follow a coherent technical logic and thus helps ensure the technical plausibility of the recommendation results. Since this dimension functions as a structural validation mechanism rather than a direct relevance indicator, β₂ is set to 0.3.

Patent element coverage evaluates whether the recommended patents contain a relatively complete set of relevant technical elements, ensuring the completeness of the recommendation results. Compared with semantic relevance and structural rationality, this metric mainly serves as a supporting indicator. Therefore, β₃ is set to 0.2.

Finally, the patent with the highest comprehensive score is selected as the recommendation result:

$$p^{\ast }=\arg \underset{p_j\in T}{\max }\mathrm{Score}(p_j)$$

(13)

Simultaneously, the algorithm provides comprehensive matching evidence, including matching scores across all dimensions and comprehensive score calculations, ensuring interpretability and traceability of recommendation results.

In summary, PR-CTD achieves supply–demand matching from complete demands to precise patent recommendations by integrating the semantic expansion capability of demand lexical trees, the deep semantic understanding capability of BERT, and the structural modeling capability of TransE, forming a precision recommendation algorithm oriented toward complete demands.

3.3. Fuzzy Recommendation Algorithm for Incomplete Demands

This study defines the recommendation algorithm for constructing multi-candidate patent sets for incomplete demands as the “Fuzzy Recommendation Algorithm for Incomplete Technological Demands” (FR-ITD). Enterprise incomplete demands suffer from ambiguous demand expression and element deficiency, making it difficult to achieve accurate patent supply–demand matching relying solely on enterprise demand texts. Therefore, the core objective of the FR-ITD algorithm is to identify multiple patent supply solutions based on elements and demand categories of incomplete demands, providing demand parties with multiple potentially needed patents for comparative selection. FR-ITD emphasizes three aspects: (1) content elements of recommended patents must match incomplete demand elements to ensure supply–demand content relevance; (2) FR-ITD must reasonably expand the recommendation scope based on demand categories to provide enterprises with potentially needed related patents; (3) the recommended patent set should present diverse technical approaches, providing enterprises with multiple options for comparative selection.

Existing patent recommendation methods exhibit certain limitations when addressing incomplete demands, primarily manifested in three aspects: First, the demand element deficiency problem—incomplete demands provide only partial element information, and matching based solely on this information tends to introduce numerous irrelevant patents due to excessively broad demand scope. Second, the demand category positioning problem—existing patent recommendation methods do not consider categories of incomplete demands, overlooking enterprises’ technology directions, technology foundations, and capability requirements, making it difficult to grasp which patents enterprises truly need. Third, insufficient diversity of recommendation solutions—most recommendation algorithms target precise matching, neglecting the actual need for diversified recommendation solutions for incomplete demands.

To address these issues, this study designs a fuzzy recommendation algorithm (FR-ITD) integrating the patent supply knowledge graph, demand categories, and LLMs. This algorithm supplements enterprise demand information through demand categories, constructing an augmented query context integrating demand element and demand category information. Patent retrieval is executed in the patent supply knowledge graph, with LLMs utilized to evaluate and screen candidate patents. Finally, comprehensive consideration of supply–demand content conformity, demand category alignment, and patent supply solution type produces patent recommendation solutions with high supply–demand matching and diverse solutions. Compared with traditional patent recommendation methods, FR-ITD offers two advantages:

(1) Demand categories supplement demand information. The demand category matrix integrates enterprise technology foundation, technology direction, and capability requirement information, reasonably expanding demand categories and refining demand bases for patent matching, which is conducive to improving relevance and effectiveness of patent recommendation results.

(2) LLMs enhance recommendation effectiveness and efficiency. FR-ITD inputs both demand information and patent information into LLMs, enabling them to identify content associations between patent solutions and incomplete demands based on semantic understanding capabilities. This guides comprehensive judgment from supply–demand content and demand category dimensions, rapidly locating patents highly relevant to incomplete demands among large-scale candidate patents, facilitating improved patent recommendation efficiency.

The FR-ITD scheme is illustrated in Figure 4. The algorithm comprises four main stages:

Stage 1: Priority Ranking of Demand Category Terms

To ensure that demand category terms highly relevant to incomplete demands are prioritized in the patent supply–demand matching process, this study conducts priority ranking of demand category terms based on the demand category term matrix and similarity between “demand category terms—demand elements”.

First, the comprehensive fitness score of each demand category term $k_i$ in the demand matrix is calculated, which is the average of similarity scores with other category terms, used to measure the association level with other category terms. A higher comprehensive fitness score indicates greater representativeness and importance in the demand, as shown in Formula (14):

$$W_{matrix}(k_i)={\displaystyle \frac{1}{\left|C\right|}}\sum _{j=1}^{\left|C\right|}{w}_{i,j}$$

(14)

where $w_{i,j}$ represents the semantic similarity between demand category term $k_i$ and the $j$-th category term from a different axis, and $C$ is the number of terms from different axes relative to demand category term $k_i$. Based on the comprehensive fitness scores, the top 10 ranked demand keywords are selected to construct a candidate vocabulary pool.

Subsequently, the semantic similarity between each candidate term and demand elements is calculated. Element information is extracted from the demand, the BERT model generates element vector ${\mathbf{v}}_{element}$, and the semantic similarity between candidate keyword $k_i$ and demand elements is calculated, as shown in Formula (15):

$$S_{base}(element,k_i)=\cos (v_{element},v_{k_i})$$

(15)

where ${\mathbf{v}}_{k_i}$ represents the BERT vector representation of keyword $k_i$. Demand category terms are ranked according to semantic similarity scores, and the top 5 terms with the highest relevance to the current demand are selected as priority demand category terms for subsequent retrieval, as shown in Formula (16):

$$Top5=\left\{k_i\right|k_i\in CandidatePool\wedge S_{semantic}(element,k_i)\in \mathrm{Top}5 \mathrm{highest} \mathrm{scores}\}$$

(16)

Stage 2: Augmented Query Context Construction

Augmented query context refers to integrating multiple types of information to form semantically more complete query content. FR-ITD organizes and combines enterprise demand texts, demand element information, and demand category terms to construct the augmented query context. This stage hierarchically organizes different types of information, providing retrieval conditions for the knowledge graph on one hand and supporting LLMs in understanding demands on the other, as shown in Formula (17):

$$Contex{t}_{enlanced}=Concat(D_{original},Se{p}_1,E_{limited},Se{p}_2,Top5,Se{p}_3,Contex{t}_{background})$$

(17)

where $D_{original}$ represents the original text of the enterprise’s incomplete demand, preserving complete semantic information of the demand; $E_{limited}$ denotes element information extracted from the demand, ensuring accuracy of patent recommendation direction; $Top5$ represents the top 5 demand category terms from Stage 1, supplementing demand categories; $Contex{t}_{background}$ comprises enterprise capability requirement terms and technology direction terms from the demand category matrix, providing enterprise background information for patent retrieval; and $Se{p}_1$, $Se{p}_2$, and $Se{p}_3$ are semantic separators ensuring that LLMs can accurately distinguish different types of input information.

Stage 3: Intelligent Retrieval and Screening Based on Knowledge Graph

Based on the augmented query context constructed in Stage 2, this stage executes intelligent retrieval and screening in the patent supply knowledge graph. This stage comprises three steps:

(1) Patent Retrieval Term Extraction

Patent retrieval terms are extracted from the augmented query context $Contex{t}_{enhanced}$, including demand elements, Top5 demand category terms, enterprise capability requirement terms, and technology direction terms. These terms are integrated to form the patent supply knowledge graph query term set.

(2) Knowledge Graph Constraint Retrieval

Based on the query term set, composite retrieval conditions are constructed in the patent supply knowledge graph. Priority is given to matching patents containing demand elements while expanding to patents associated with demand category terms. Entity retrieval is executed through MATCH pattern matching and WHERE condition filtering to obtain the initial candidate patent set $R$.

(3) LLM-Based Candidate Patent Screening

The initial candidate patent set $R$ and its element information, together with the augmented query context, are jointly input into the LLM. A prompt is designed to guide the LLM in executing supply–demand matching analysis. The LLM’s candidate patent screening results are denoted as $C_{initial}$, serving as patents to be scored in Stage 4.

Stage 4: Patent Recommendation Scoring

Stage 3 obtained the candidate patent set $C_{initial}$. To evaluate the matching degree of candidate patents, FR-ITD quantitatively assesses the matching degree of candidate patents from two dimensions: supply–demand content conformity and demand category alignment.

(1) Supply–Demand Content Conformity (For Incomplete Demands)

Supply–demand content conformity measures the degree to which candidate patents respond to identified elements in incomplete demands. For each patent $p_i$ in the candidate patent set, its supply–demand content conformity is calculated as shown in Formula (18):

$$S_{element}(p_i)={\displaystyle \frac{1}{\mid E_{limited}\mid }}\sum _{e\in E_{limind}}\underset{c^{\prime }\subset \mathrm{Elcmcnts}(p_i)}{\max }\cos (v_e,v_{e^e})$$

(18)

where $E_{limited}$ represents the demand element set, and $Elements\left(p_i\right)$ represents the element set contained in patent $p_i$. ${\mathbf{v}}_e$ and ${\mathbf{v}}_{e^{\prime }}$ denote the BERT vector representations of demand elements and patent elements, respectively. By calculating similarity between demand elements and patent elements, this metric ensures that each known demand element can find corresponding responsive content in candidate patents, preventing recommendation results from deviating from demands.

(2) Demand Category Alignment

Demand category alignment is used to evaluate the degree of semantic matching between candidate patents and Top5 demand category terms, reflecting the response level of patent content to enterprise technological demand categories, as shown in Formula (19):

$$S_{category}(p_i)={\displaystyle \frac{1}{\left|Top5\right|}}\sum _{k\in Top5}\underset{e^{\prime }\in \mathrm{Elements}(p_i)}{\max }\cos (v_k,v_{e^c})$$

(19)

where $Top5$ represents the Top5 demand category terms from Stage 1, and ${\mathbf{v}}_k$ denotes the vector representation of Top5 demand category terms. This metric measures the matching degree between patent content and demand categories by calculating semantic similarity between Top5 demand category terms and patent element vectors, enabling the algorithm to prioritize recommending patents that satisfy enterprise demand categories.

The final comprehensive score calculation is shown in Formula (20):

$$Scor{e}_{final}(p_i)={\gamma }_1\cdot S_{element}(p_i)+{\gamma }_2\cdot S_{category}(p_i)$$

(20)

γ₁ represents the weight parameter for supply–demand content conformity, while γ₂ represents the weight parameter for demand category alignment. In this study, γ₁ is set to 0.4 and γ₂ is set to 0.6. Supply–demand content conformity measures the semantic correspondence between enterprise technological demands and patent technical content. It serves as a fundamental indicator for establishing supply–demand relationships in the patent recommendation process. The technical elements identified from demand texts represent the core content of enterprise technological needs. By semantically matching these elements with patent technical information, it is possible to determine whether a patent has the potential to address the technological problem described in the demand. Therefore, a moderate weight is assigned to supply–demand content conformity to ensure that the recommendation results maintain basic technical relevance while leaving sufficient space for demand category expansion. Demand category alignment evaluates the consistency between patent technologies and enterprise technological demands at the level of technological direction and capability foundation. Demand categories are typically constructed based on information such as the enterprise’s technological field, capability requirements, and technology development direction, which together reflect the overall technological positioning of the enterprise in the innovation process. In the patent recommendation framework, demand category alignment helps filter technical solutions that are consistent with the enterprise’s technological foundation and ensures that the recommended patents remain aligned with the enterprise’s technological development direction. Therefore, a relatively higher weight is assigned to this dimension, with γ₂ = 0.6. Overall, the above weight settings are determined according to the functional roles of different evaluation dimensions in the patent recommendation process, ensuring the stability and rationality of the recommendation results.

Patents are ranked in descending order according to comprehensive scores. The method ultimately outputs a list of the top five recommended patents, with each patent accompanied by its supply–demand content conformity score and demand category alignment score, providing demand parties with multiple patent recommendation results with interpretable matching evidence.

FR-ITD addresses the element deficiency problem of incomplete demands through a four-stage cascading mechanism of “demand category term priority ranking + augmented context construction + knowledge graph intelligent retrieval and screening + two-dimensional scoring,” integrating demand category information to achieve supply–demand matching from vague demands to diversified patent candidate sets.

4. Experiments

To validate the effectiveness of the supply–demand matching and patent recommendation framework proposed in this study, the lithium battery technology domain is selected as the empirical context. A validation dataset is constructed comprising real enterprise technology demands and the corresponding patent transfer records that occurred after those demands were published. By comparing patent recommendation results against actual patent transfer outcomes, the applicability and effectiveness of PR-CTD (Precise Recommendation for Complete Technology Demands) and FR-ITD (Fuzzy Recommendation for Incomplete Technology Demands) are evaluated under real-world application scenarios.

4.1. Construction of the Empirical Dataset

(1) Acquisition of enterprise technology demand data

Enterprise technology demand data were retrieved from the public science and technology achievement transfer and transformation service platforms of 12 provinces, namely Anhui, Fujian, Gansu, Hebei, Henan, Heilongjiang, Hunan, Jilin, Jiangxi, Shanxi, Shaanxi, and Yunnan. The search was conducted using keywords including “lithium battery,” “power battery,” “energy storage battery,” “solid-state battery,” “battery materials,” “battery manufacturing,” and “battery recycling,” covering the period from 1 January 2015 to 31 December 2024. Data fields including enterprise name, demand content, and publication date were collected via Python 3.8 web crawlers, yielding an initial sample of 233 demand entries. After manual screening, enterprise status verification, and deduplication, a final set of 197 enterprise technology demands meeting the research criteria was obtained, involving 139 enterprises.

(2) Acquisition and screening of patent transfer data

Patent transfer data were sourced primarily from the PatSnap patent database. Using the names of demand-submitting enterprises as the assignee or licensee field, patent transfer records occurring after the publication of each corresponding demand were retrieved. Initial search results indicated that 62 enterprises had collectively engaged in 131 patent transfer transactions following the submission of their demands. After merging patent family duplicates and excluding intra-enterprise ownership changes, 105 valid patent transfer records were retained.

(3) Construction of the demand-driven patent transfer validation dataset

Among the 105 candidate records, the temporal ordering of events indicates only that patent transfer activities could logically be associated with enterprise technology demands. Further identification of transfer cases that were genuinely driven by those demands was therefore required. This study employed a BERT pre-trained language model to encode enterprise technology demand texts and transfer patent abstract texts into semantic vectors. Cosine similarity was used to measure supply–demand relevance, as defined in Equation (21):

$$Sim(d,p)={\displaystyle \frac{v_d\cdot v_p}{\parallel v_d\parallel \parallel v_p\parallel }}$$

(21)

This metric reflects the degree of content-level relevance between the supply and demand sides. A threshold was set to preliminarily exclude patent transfer records with similarity scores below that value. For records with similarity scores near the threshold boundary, manual verification was conducted to determine whether each case constituted a demand-driven patent transfer.

Integrating the results of semantic similarity computation and manual assessment, 25 patent transfer instances were ultimately confirmed as reasonably attributable to enterprise technology demands. These transfers involve 22 technology demands from 19 enterprises and constitute the demand-driven patent transfer validation dataset used for empirical evaluation. The corresponding supply–demand relevance scores and relevance classification results are presented in

Table 4. Validation dataset of demand-driven patent transfers.

ID	Enterprise Name	Enterprise Demand Content	Transferred Patent Number	Original Patent Holder	Supply–Demand Relationship	Supply–Demand Relevance Score
1	Jiujiang ** New Materials Co., Ltd.	Technical support for lithium battery separator technology. The company’s primary product is a wet-process lithium battery separator. The enterprise adopts a wet-process dual-layer structure with one-piece forming during membrane fabrication, which differs from the dry-process multi-layer lamination approach. This method offers advantages including high yield and excellent uniformity. Expert technical support in lithium battery separator technology is required.	CN103413907B	Shenzhen ** New Materials Co., Ltd.	The demand side is a wholly owned subsidiary of the supply side.	0.7266
2	Jiangxi ** Recycling Technology Co., Ltd.	Efficient and clean full-component recovery and utilization technology and industrialization for retired lithium iron phosphate batteries. The project requires the development of: lithium-phosphorus-iron separation technology, iron-phosphorus precipitation and purification technology, anhydrous iron phosphate preparation technology, and lithium salt preparation technology from lithium-containing solutions (e.g., lithium carbonate). Target indicators: comprehensive lithium recovery rate ≥95%; iron and phosphorus recovery rate ≥92%; lithium carbonate product meeting battery-grade lithium carbonate standard (YS/T 582-2013) [46]; iron phosphate product meeting battery-grade standard (HG/T 4701-2014) with titanium and aluminum impurity content below 10 ppm [47].	CN104925837B	Jiangxi ** Lithium Industry Group Co., Ltd.	The demand side is a wholly owned subsidiary of the supply side.	0.8165
3	Jiangxi ** Recycling Technology Co., Ltd.	Intelligent disassembly equipment for retired lithium battery jellyroll cores. The objective is to jointly develop intelligent disassembly equipment for retired lithium battery jellyroll cores (unfilled with electrolyte) to meet the development needs of capacity expansion, labor reduction, and improved working conditions.	CN114940296B	Jiangxi ** Technology Innovation Research Institute	The supply and demand parties have previously collaborated in the field of intelligent disassembly technology for retired lithium batteries.	0.6529
4	Jiangxi ** Recycling Technology Co., Ltd.	Intelligent and precise disassembly of spent lithium power batteries. The enterprise has established mature processes for spent battery recycling, including the recovery of ternary lithium batteries to produce ternary precursor materials and the recovery of spent lithium iron phosphate batteries to produce lithium carbonate. The primary challenge is the low level of automation in battery pack and cell disassembly, which currently relies mainly on manual labor with mechanical assistance. The goal is to develop a complete set of intelligent and precise disassembly equipment for battery packs and cells, capable of flexible loading and unloading, intelligent identification, material control, and intelligent resource scheduling for lithium batteries of various specifications. Requirements: compatibility with no fewer than 20 battery specifications, models, and sizes; complete pack disassembly efficiency of no less than 10 units per hour; individual cell disassembly efficiency of no less than 360 units per hour. The equipment must incorporate safety, corrosion resistance, fire protection, and remote monitoring visualization features.	CN114940296B	Jiangxi ** Technology Innovation Research Institute	The supply and demand parties have previously collaborated in the field of intelligent disassembly technology for retired lithium batteries.	0.7976
5	Jiangxi ** New Energy Technology Co., Ltd.	Key technologies for resource recovery and lithium extraction from retired lithium batteries. Application context: retired lithium battery recycling and reuse. Key technologies include efficient electrode material stripping and short-process selective lithium extraction from retired power lithium batteries. Target indicators: electrode material recovery rate after pyrolysis-crushing-separation treatment ≥98%; residual organic matter, copper, and aluminum foil content in electrode powder ≤0.5%, ≤1%, and ≤0.8%, respectively; selective separation of lithium ions from leachate with a lithium recovery rate ≥97%.	CN106191445B	Nanchang Hangkong University	The supply and demand parties have previously engaged in industry–academia–research collaboration.	0.7486
6	Yichang ** Technology Co., Ltd.	R&D of a TPMS lithium micro-power source. By introducing positive electrode additives, the project aims to optimize and improve the cathode formulation, cathode sheet fabrication process, cathode drying process, and assembly process. The target battery features high capacity and high power output, with a simple production process suitable for mass production. Compared with conventional coin-type batteries, the optimized cathode powder fabrication process incorporates positive electrode additives during the mixing stage to enhance battery capacity, while the current collector ring structure adopted for the cathode sheet enables higher power output.	CN104810489B	Wuhan ** Technology Co., Ltd.	The supply and demand parties both belong to ** Technology Group.	0.7209
7	Daqing ** Energy Technology Co., Ltd.	Cold-resistant technology for graphene lithium batteries. The goal is to achieve rapid charging of graphene lithium batteries within 12 min and a driving range of 80 km under northern cold-climate conditions.	CN209001087U	Shandong ** New Energy Technology Co., Ltd.	The supply and demand parties have no equity, holding, or other capital relationships, nor any publicly documented cooperative relationship.	0.7589
8	Fu’an ** Energy Materials Co., Ltd.	Development and industrialization of high-nickel single-crystal NCM ternary precursor materials. The enterprise seeks to establish a university-enterprise collaboration to address challenges in the development of ternary cathode materials. Domestic and international cathode material manufacturers predominantly use co-precipitation combined with high-temperature solid-state synthesis to produce cathode materials, yielding micron-scale spherical secondary particles composed of submicron-scale primary particles. With repeated charge–discharge cycling, microcracks or pulverization readily occur at the interfaces between primary particles within the secondary particles, increasing interfacial resistance and polarization. The high porosity within secondary spherical particles leads to greater contact area, more side reactions, and severe gas generation, ultimately degrading battery cycling performance and safety.	CN107331859B	Jingmen ** New Materials Co., Ltd.	The supply side is the controlling shareholder of the demand side.	0.8088
9	Ganzhou ** Technology Co., Ltd.	Construction of high-performance silicon-carbon anode materials and investigation of their lithium storage performance. Application: anode materials for lithium-ion batteries. Target metrics: energy density, rate capability, and cycling performance under high and low temperature conditions.	CN116083927B	Jiangxi University of Science and Technology	The supply and demand parties have previously engaged in industry–academia–research collaboration.	0.5089
10	Guangxi ** New Energy Technology Co., Ltd.	Development and industrialization of key manufacturing technologies for high-nickel material lithium-ion batteries. The project covers: slurry preparation and coating processes for high-nickel ternary cathodes and silicon-carbon anodes; performance evaluation and application of high-strength ultra-thin separators after ceramic coating; electrolyte formulation research; and formation and aging process development. The goal is to develop high-nickel, high-specific-energy lithium-ion batteries and achieve industrialization. Specific objectives are as follows. (1) High capacity: using high-nickel cathode materials in existing 18,650 cylindrical batteries, resolving slurry moisture absorption during processing and moisture control to meet theoretical design capacity requirements, with an approximately 15% improvement over current technology. (2) High volumetric density: using a mixture of high-capacity silicon-carbon and high-compaction graphite anode materials in existing 18,650 cylindrical batteries, addressing compaction density enhancement and excessive anode sheet expansion. (3) Safety: adopting separators with ceramic coating and improved calendering properties to ensure that electrolyte penetration efficiency is not compromised and safety risks are not increased under high-capacity, high-compaction conditions. (4) Overall performance: achieving high safety, long range, fast charge–discharge capability, and good manufacturing efficiency.	CN105336953B	Guangxi Normal University	The supply and demand parties have previously engaged in industry–academia–research collaboration.	0.8256
11	Henan ** New Energy Materials Technology Co., Ltd.	Lithium battery separator products. The lithium battery separator industry presents high technical barriers. Current R&D projects include: three-layer co-extruded PP/PP/PP homogeneous high-strength separators; three-layer co-extruded PP/PE/PP heterogeneous high-safety separators; organic-inorganic functional composite high-safety separators; aqueous polyimide high-temperature-resistant composite separators; ion-adsorbing long-life composite separators; semi-solid-state battery composite separators; and ultra-thin high-strength multilayer composite separators. Challenges encountered during development include wrinkling, surface non-uniformity, and a high density of crystalline defects. Domestically sourced raw materials also suffer from insufficient purity and precision, unstable production batches, and a limited number of qualified suppliers.	CN111697189B	Foshan ** Optoelectronic Materials Co., Ltd.	The supply and demand parties have no equity, holding, or other capital relationships, nor any publicly documented cooperative relationship.	0.7578
			CN111816826B			0.7484
			CN214203906U			0.8826
			CN107195838B			0.8899
12	Henan ** New Energy Materials Technology Co., Ltd.	Three-layer composite lithium battery separator casting technology. Standard plastic film processing equipment typically includes screen changers and quick-change screen devices, with relatively low requirements on screen change intervals. In contrast, processing equipment for lithium battery separators imposes strict requirements on both the size of crystalline defects and the number of defects tolerable per unit area, and longer intervals between screen changes are preferred to improve product yield. To avoid thermal degradation of materials during processing, the pursuit of excellence in separator manufacturing centers on the use of low-temperature, high-throughput screws, large-area long-life filter screens, and screen changers operating at appropriate pressure.	CN111697189B	Foshan ** Optoelectronic Materials Co., Ltd.	The supply and demand parties have no equity, holding, or other capital relationships, nor any publicly documented cooperative relationship.	0.6082
			CN214203906U			0.8709
13	Jixi ** New Energy Technology Co., Ltd.	Development technology for long-cycle, low-cost natural graphite anode materials. The enterprise has investigated graphite surface modification techniques to increase interlayer spacing or structural stability, and urgently requires further theoretical and technical support. Specific areas of interest include elemental doping, composite carbon material preparation, graphite surface modification, and pore-filling techniques to control the size and distribution of internal voids in spherical graphite particles, thereby achieving high compaction density and low expansion characteristics for long-cycle, low-cost graphite anodes. Technical targets: tap density ≥1.00 g/cc; specific capacity ≥355 mAh/g; 1.5C/0.5C capacity retention ≥80%; capacity retention ≥80% after 1500 cycles at 1C discharge; overall cost below 20,000 RMB per ton.	CN104269555B	BTR ** Group Co., Ltd.	The demand side is a wholly owned subsidiary of the supply side.	0.9777
			CN105261734B			0.9983
14	Jiangxi ** New Energy Technology Co., Ltd.	Design solutions for precision structural components of power batteries. Square lithium batteries are widely used in new energy vehicle power systems. The enterprise seeks to surpass current technical benchmarks to effectively improve performance indicators and reduce manufacturing costs, with the aim of obtaining proprietary intellectual property rights over the relevant technologies, either in whole or in part.	CN109301103B	Dongguan ** New Energy Technology Co., Ltd.	The supply side is a wholly owned subsidiary of the demand side.	0.7242
15	Jiangxi ** New Energy Materials Technology Co., Ltd.	Tail gas separation and recovery technology. The company is engaged in the production of lithium hexafluorophosphate and faces challenges in tail gas treatment, specifically the need to separate and purify hydrogen chloride, hydrogen fluoride, and phosphorus pentafluoride.	CN114956021B	Shanghai ** New Materials Technology Co., Ltd.	The demand side is an indirect wholly owned subsidiary of the supply side.	0.5564
			CN113353958B			0.7333
16	Jiangxi ** Lithium-Rich Technology Co., Ltd.	Research on cycle life improvement and application performance of lithium-rich manganese-based cathode materials. The research is intended to be conducted within the factory production environment and covers: (1) benchmarking material performance and investigating blending systems, with a focus on analyzing key characteristics of the material under medium-to-high voltage conditions; and (2) investigating coating strategies for the material, evaluating the effect of process parameters on electrochemical performance, conducting full-cell prototyping and performance evaluation, and improving the high-temperature cycling life of full cells.	CN114256457B	Guolian ** Battery Research Institute Co., Ltd.	The supply and demand parties have no equity investment relationship but have jointly filed patent applications.	0.8167
17	Jiangxi ** Intelligent Technology Co., Ltd.	R&D and industrialization of 4.45 V fast-charging lithium-ion batteries. The enterprise requires theoretical research and material system development for high-voltage fast-charging lithium-ion batteries, along with access to existing R&D and pilot-scale production facilities.	CN112599782A	Phoenix ** Technology Co., Ltd.	The supply and demand parties have no direct equity control relationship; both belong to the Shenzhen ** Battery Co., Ltd. corporate system.	0.5862
18	Jiangxi ** New Energy Co., Ltd.	R&D of lithium-ion batteries and investigation of materials required for lithium-ion battery manufacturing.	CN218677247U	Guo * (individual inventor)	The supply and demand parties have no identifiable association.	0.5076
19	Jiangxi ** New Energy Technology Co., Ltd.	Automated disassembly technology for spent lithium batteries and high-technology-content precursor preparation technology. The demand covers: technologies for efficient recovery of valuable metals from spent lithium batteries; integration of informatization and industrialization; and high-technology-content precursor preparation technology to achieve efficient and automated production of ternary cathode materials with high specific capacity, long cycle life, and low self-discharge rate.	CN108940428B	Li * (individual inventor)	The supply and demand parties have no identifiable association.	0.6128
			CN108963376B			0.5343
20	Jiujiang ** High-Tech Materials Co., Ltd.	High-voltage electrolyte for lithium-ion batteries. Technical requirements for the electrolyte: moisture content ≤25 ppm; free acid ≤50 ppm; sulfate ≤5 ppm; chloride ≤1 ppm; individual metal impurity ≤1 ppm. The electrolyte must be suitable for use in ternary cathode material lithium-ion batteries operating within a 3–4.35 V charge–discharge window, with the process index improved to 70%.	CN119695238B	Guangzhou ** High-Tech Materials Co., Ltd.	The demand side is a wholly owned subsidiary of the supply side.	0.7454
			CN119674234B			0.6308
			CN113745660B			0.7411
			CN113956282B			0.7645
21	Lingbao ** Electronic Technology Co., Ltd.	Research on 6 μm double-bright-surface ultra-high tensile strength copper foil for lithium batteries. The existing 6 μm double-bright-surface ultra-high tensile strength copper foil for lithium batteries has a high-temperature tensile strength above 400 MPa but below 500 MPa. The project aims to increase tensile strength by optimizing the ratio of currently used additives or by introducing new additives, with specific hard and optional target indicators to be achieved.	CN106350836B	Nanjing ** Electronic Technology Co., Ltd.	The supply and demand parties have no direct cross-shareholding relationship; both belong to Longdian ** Holding Group Co., Ltd.	0.7218
22	Xinxiang ** Technology Co., Ltd.	Polyolefin microporous separators. To further upgrade its products, the enterprise aims to overcome the key technical challenge of ensuring the safe application of next-generation polyolefin microporous separators for lithium batteries.	CN102290549B	Xinxiang ** New Energy Materials Co., Ltd.	The supply side is a wholly owned subsidiary of the demand side.	0.5990
			CN102267229B			0.6991

* To protect the business privacy of the enterprises involved, company names and patent assignee names in this table have been partially anonymized, with brand-specific keywords replaced by “**”. This does not affect the authenticity or integrity of the data.

.

In terms of overall supply–demand relevance, the mean score across all samples is approximately 0.71, with the majority of samples falling in the range of 0.65 to 0.90. Approximately 60% of the patent transfer samples achieve scores above 0.75. Regarding the relationship between supply and demand parties, the dataset encompasses five relationship types: parent-subsidiary companies, industry–academia–research collaborations, entities within the same holding structure, co-patent applicants, and parties with no identifiable prior association. These cases span diverse patent transfer contexts including intra-group consolidation, industry–academia–research commercialization, and open market transactions, thereby ensuring the objectivity and validity of the subsequent empirical evaluation of both recommendation algorithms.

4.2. Identification of Enterprise Technology Demands

This section performs element extraction on enterprise technology demands in the lithium battery domain, identifying the five element types involved in each demand: material, method, efficacy, product, and application. The identification results are presented in

Table 5. Results of enterprise technological demand element identification in patent transfer validation dataset.

Demand ID	Material	Method	Efficacy	Product	Application
1	—	① Wet-process dual-layer structure; ② One-piece forming membrane fabrication	① High yield; ② Excellent uniformity	① Wet-process lithium battery separators; wet-process separator production line	① Lithium battery separators
2	① Retired lithium iron phosphate batteries	① Lithium-phosphorus-iron separation technology; ② Iron-phosphorus precipitation and purification technology; ③ Anhydrous iron phosphate preparation technology; ④ Lithium salt preparation technology from lithium-containing solutions (e.g., lithium carbonate)	① Comprehensive lithium recovery rate ≥95%; ② Iron and phosphorus recovery rate ≥92%; ③ Lithium carbonate product meeting battery-grade lithium carbonate standard; ④ Iron phosphate product meeting battery-grade standard with titanium and aluminum impurity content below 10 ppm	① Lithium carbonate; ② Iron phosphate	① Full-component recovery and utilization of retired lithium iron phosphate batteries
3	① Retired lithium battery jellyroll cores (unfilled with electrolyte)	① Intelligent disassembly	① Capacity expansion; ② Labor reduction; ③ Improved working conditions	① Complete set of disassembly equipment for retired lithium battery jellyroll cores	① Disassembly of retired lithium battery jellyroll cores
4	① Spent batteries; ② Ternary lithium batteries; ③ Spent lithium iron phosphate batteries	① Intelligent and precise disassembly; ② Flexible loading and unloading; ③ Intelligent identification; ④ Material control and intelligent resource scheduling	① Compatibility with no fewer than 20 battery specifications, models, and sizes; ② Complete pack disassembly efficiency ≥10 units/hour; ③ Individual cell disassembly efficiency ≥360 units/hour; ④ Safety, corrosion resistance, fire protection, and remote monitoring visualization	① Intelligent and precise disassembly equipment for battery packs and cells; ② Ternary precursor materials; ③ Lithium carbonate	① Disassembly and recovery of spent lithium power batteries
5	① Retired power lithium batteries; ② Electrode materials	① Pyrolysis-crushing-separation; ② Short-process selective lithium extraction	① Electrode material recovery rate ≥98%; ② Residual organic matter in electrode powder ≤0.5%; ③ Copper ≤1%; ④ Aluminum foil ≤0.8%; ⑤ Lithium recovery rate ≥97%	① Lithium	① Recovery and reuse of retired lithium batteries
6	① Cathode additives	① Cathode formulation optimization; cathode sheet fabrication process optimization; ② Cathode drying process optimization; ③ Assembly process optimization; ④ Current collector ring structure	① Room-temperature discharge capacity ≥200 mAh; ② Discharge capacity at low temperature (−40 °C) ≥120 mAh; ③ Discharge capacity at high temperature (105 °C) ≥200 mAh; ④ Normal operation after continuous storage at 125 °C for 100 h; ⑤ Normal operation after continuous storage at 150 °C for 24 h	① CR2032HT battery; ② TPMS lithium micro-power source	① TPMS
7	① Graphene	—	① Fast charging within 12 min; ② Driving range of 80 km; ③ Cold-weather resistance	① Graphene lithium battery	① Northern cold-climate environments
8	① High-nickel materials with nickel content of 83–90%; ② High-nickel core–shell concentration gradient materials; ③ Multi-element doped materials; ④ High-nickel cobalt-free materials	① Co-precipitation combined with high-temperature solid-state synthesis; ② Co-precipitation method	① Good crystallinity; ② Structural stability; ③ Prevention of microcracks or pulverization; ④ Reduced interfacial resistance; ⑤ Reduced polarization; ⑥ Reduced side reactions and gas generation	① High-nickel single-crystal NCM ternary precursor materials; ② Ternary cathode materials	① Lithium battery cathode materials
9	① Silicon-carbon anode materials	—	① High energy density; ② High rate capability; ③ Good cycling performance under high and low temperature conditions	① Silicon-carbon anode materials	① Anode materials for lithium-ion batteries
10	① High-nickel ternary cathode materials; ② Silicon-carbon anode materials; ③ High-strength ultra-thin separators; ④ Ceramic-coated separators; ⑤ Electrolyte; ⑥ High-compaction graphite	① Slurry preparation and coating processes; ② Formation and aging processes; ③ Ceramic coating; ④ Calendering	① Capacity approximately 15% higher than current technology; ② High volumetric density; ③ High safety; ④ Long range; ⑤ Fast charge and discharge; ⑥ Good manufacturing efficiency	① High-nickel, high-specific-energy lithium-ion batteries; ② 18,650 cylindrical lithium-ion batteries	① Lithium-ion battery manufacturing
11	① Polypropylene raw materials	① Three-layer co-extrusion	① High strength; ② High safety; ③ High temperature resistance; ④ Long service life; ⑤ Ultra-thin profile	① PP/PP/PP homogeneous high-strength separators; ② PP/PE/PP heterogeneous high-safety separators; ③ Organic-inorganic functional composite high-safety separators; ④ Aqueous polyimide high-temperature-resistant composite separators; ⑤ Ion-adsorbing long-life composite separators; ⑥ Semi-solid-state battery composite separators; ⑦ Ultra-thin high-strength multilayer composite separators	① Lithium battery separators
12	—	① Screen changer and quick-change screen devices; ② Low-temperature high-throughput screws; ③ Large-area long-life filter screens	① Extended screen change intervals; ② Prevention of thermal degradation of materials; ③ Improved product yield	① Three-layer composite lithium battery separators	① Lithium battery separator manufacturing
13	① Natural graphite and composite carbon materials	① Surface modification technology; ② Elemental doping; ③ Graphite pore-filling technology	① Tap density ≥1.00 g/cc; ② Specific capacity ≥355 mAh/g; ③ 1.5C/0.5C capacity retention ≥80%; ④ Capacity retention ≥80% after 1500 cycles at 1C discharge; ⑤ Overall cost below 20,000 RMB/ton; ⑥ High compaction density and low expansion	① Spherical graphite; ② Natural graphite anode materials	① Lithium battery anode materials
14	—	① Structural design solutions; ② Engineering implementation technology; ③ Scale manufacturing	① Improved performance indicators; ② Reduced manufacturing costs	① Precision structural components for power batteries; ② Prismatic lithium battery cells	① Power batteries for new energy vehicles
15	① Hydrogen chloride; ② Hydrogen fluoride; ③ Phosphorus pentafluoride	① Separation and purification	—	① Lithium hexafluorophosphate	① Lithium hexafluorophosphate production
16	① Lithium-rich manganese-based materials	① Material performance benchmarking; ② Blending system investigation; ③ Coating research	① Improved high-temperature cycling life of full cells	① Lithium-rich manganese-based cathode materials; ② Full cells	① Medium-to-high voltage application scenarios
17	① High-voltage fast-charging material systems	—	① 4.45 V high voltage; ② Fast charging	① 4.45 V fast-charging lithium-ion batteries	① Fast-charging lithium-ion battery applications
18	① Materials required for lithium-ion battery manufacturing	—	—	① Lithium-ion batteries	—
19	① Spent lithium batteries	① Automated disassembly technology; ② Precursor preparation technology; ③ Integration of informatization and industrialization	① Efficient recovery of valuable metals; high specific capacity; ② Long cycle life; ③ Low self-discharge rate; ④ Efficient and automated production	① Ternary cathode material precursors	① Spent lithium battery recycling
20	① Electrolyte	—	① Moisture content ≤25 ppm; ② Free acid ≤50 ppm; ③ Sulfate ≤5 ppm; ④ Chloride ≤1 ppm; ⑤ Individual metal impurity ≤1 ppm; ⑥ Suitable for 3–4.35 V charge–discharge operation; ⑦ Process index improved to 70%	① High-voltage electrolyte	① Ternary cathode material lithium-ion batteries
21	① Additives; ② Copper foil	① Optimization of additive ratios; ② Introduction of new additives	① High-temperature tensile strength increased to above 500 MPa	① 6 μm double-bright-surface ultra-high tensile strength copper foil for lithium batteries	① New energy vehicle manufacturing
22	① Polyolefin	—	① Breakthrough in safe application performance	① Polyolefin microporous separators	① Next-generation lithium batteries

‘—’ indicates the corresponding element is absent in the demand and the content is empty

. Among the 22 enterprise technology demands identified in the lithium battery domain, 12 are classified as complete demands and 10 as incomplete demands. This result indicates that enterprise technology demands vary substantially in their level of expression completeness, with nearly half of the demands exhibiting insufficient information or imprecise articulation, which indirectly underscores the necessity of developing differentiated patent recommendation algorithms. For the 10 identified incomplete demands (Demand IDs 1, 7, 9, 12, 14, 15, 17, 18, 20, and 22), this study performs categorical expansion along three dimensions: enterprise industry background, historical technology development, and technology capability requirements. The supplementary results are presented in

Table 6. Results of category supplementation for incomplete technological demands in the patent transfer validation dataset.

Demand ID	Technology Direction	Technology Base	Capability Requirements
1	Nano-scale microporous membranes; wet-process lithium battery separators; lithium batteries; power batteries; 3C batteries	Wet-process thin films; wet-process separators; simultaneous stretching machines; quench cooling systems; laser slitting; thin film coating; ceramic separators; alumina coating layers; microporous structures	—
7	Battery manufacturing	Protocol control; charging pile protection; adjustable charging piles; solar panels; photovoltaic power generation for tower base stations; battery sorting and recycling; battery management systems (BMSs); graphene batteries; graphene processing; ultrasonic dispersion	—
9	Battery manufacturing	Porous silicon anode materials; silicon–carbon anode materials; silicon monoxide; plasma etching; CVD vapor-phase coating; in situ deposition of conductive carbon layers; amorphous carbon layers; hard carbon coating; vapor-phase deposition of silane; cycling performance	Mechanical engineering; chemical engineering
12	Lithium battery separator production	Polyolefin microporous base membranes; multilayer microporous membranes; heat-resistant coatings; ceramic coating; thermal safety; thermal shrinkage; ion conduction; separator mechanical properties and stiffness; surface grafting modification; quality inspection	—
14	Battery component manufacturing; battery component sales	Thermal runaway protection; cell heat dissipation and cooling channels; electrolyte replenishment; electrolyte sealing and leak prevention; internal short-circuit risk control; overcurrent capability enhancement; pressure relief and explosion prevention; online temperature-pressure monitoring; high-energy-density thin-profile design; high-voltage insulation and sealing	Component selection and design
15	New material technology R&D	Lithium hexafluorophosphate precursor preparation; continuous production processes for hexafluorophosphate salts; purification of battery-grade hexafluorophosphate; key lithium salt raw materials for electrolytes; clean production processes; uniform mixing control of reaction systems; gas–liquid reaction mass transfer enhancement; preparation of high-purity phosphorus pentachloride feedstock; anti-caking storage and transportation; moisture control; safety and stability of lithium battery electrolytes	Biochemical product technology R&D; new material technology R&D; chemical product manufacturing
17	Battery manufacturing	Electrochemical activity restoration of cathode materials; lithium manganese iron phosphate; high-capacity anode materials; graphene composite electrode materials; electrode carbon coating and surface modification; electrode sheet coating tension control; tab ultrasonic welding; jellyroll structure optimization; cell casing; battery electrolyte injection; electrolyte stability control; fast charging	Lithium battery material development
18	Lithium-ion cells; lithium-ion batteries; power batteries; energy storage batteries; battery materials; mobile power sources; power battery management systems (BMSs); power management systems	Integrated electrolyte injection and formation processes; pressure-controlled formation; electrode sheet coating systems; electrode sheet baking and drying; battery slurry preparation; short-circuit cell detection; rapid battery testing; cell structural design; cathode coating modification; battery safety	Evaluation of cathode and anode materials, separators, and electrolytes; design, fabrication, and production of lithium-ion battery cells
20	Lithium-ion batteries	Lithium fluoride purification; lithium hexafluorophosphate synthesis; lithium bis(fluorosulfonyl)imide; lithium difluoro(oxalato)borate; lithium difluorophosphate; fluoroethylene carbonate; electrolyte additive formulation; SEI/CEI film repair; lithium iron phosphate cathode regeneration and recovery; microfluidic continuous preparation	Electrolyte additive R&D; lithium-ion battery material R&D; lithium hexafluorophosphate production
22	R&D of specialized separator materials for lithium-ion batteries	Five-layer co-extruded dry-process separators; semi-solid-state composite separators; nano-elastomer thermal shutdown separators; high-rate functional composite membranes; high-voltage-resistant nano-coated separators; POSS-reinforced aramid-coated separators; nano-plasticized porous PE dry-stretch membranes; PP microporous membrane porosity regulators; separator deposition layer generation devices; online separator porosity testing systems	Battery separator production

‘—’ indicates that the corresponding capability requirement information is absent for this enterprise, and the content is empty.

.

4.3. Construction of the Patent Supply Knowledge Graph

This section applies the patent supply knowledge graph construction method described in Section 2.2 to build a corresponding patent supply knowledge graph for each of the 22 enterprise technology demands. An entity extraction approach combining structured field extraction and large language model recognition is employed to extract five types of element entities: material, method, efficacy, product, and application. Relation triples are then generated using a rule-driven relation construction method, and the resulting entities and relations are imported into the Neo4j graph database. The scale statistics of the patent supply knowledge graph corresponding to each demand are presented in

Table 7. Statistics of patent retrieval, selection, and patent supply knowledge graph scale corresponding to enterprise technological demands.

Demand ID	Demand Type	Initial Patent Count	Patent Count After Screening	Patent Supply Knowledge Graph
				Entity Count	Relation Count
1	Incomplete Demand	658	525	6439	14,297
2	Complete Demand	910	726	6962	20,065
3	Complete Demand	335	265	3272	7287
4	Complete Demand	1889	1510	13,513	41,027
5	Complete Demand	4860	3887	31,141	105,734
6	Complete Demand	735	587	6692	15,480
7	Incomplete Demand	478	381	4310	11,052
8	Complete Demand	920	733	6398	20,531
9	Incomplete Demand	1105	883	7947	26,457
10	Complete Demand	1279	1022	8875	27,160
11	Complete Demand	1013	810	8602	23,138
12	Incomplete Demand	728	583	7238	14,327
13	Complete Demand	246	196	2266	5632
14	Incomplete Demand	3260	2608	27,300	36,536
15	Incomplete Demand	535	426	4242	10,837
16	Complete Demand	815	651	5564	18,597
17	Incomplete Demand	1625	1300	11,561	35,340
18	Incomplete Demand	755	604	6235	15,068
19	Complete Demand	795	636	6577	17,557
20	Incomplete Demand	237	188	2262	4749
21	Complete Demand	189	150	1653	3807
22	Incomplete Demand	1568	1253	10,738	35,556

. Collectively, the 22 knowledge graphs contain 189,787 entity nodes and 510,234 relations, providing a sufficiently large and structurally complete patent supply-side data foundation for the subsequent recommendation algorithms.

4.4. Empirical Evaluation of the Precise Recommendation Algorithm for Complete Demands

This section uses supply–demand content conformity as the core evaluation metric, computing the degree of supply–demand matching between recommended patents and enterprise demands across five element dimensions: material, method, efficacy, product, and application.

Table 8. Supply–demand content matching degree of PR-CTD recommended patents and ranking of actually commercialized patents.

Demand ID	Patent	Supply–Demand Content Conformity					Actual Transferred Patent Ranking
		Material	Method	Efficacy	Product	Application
2	RP: CN114988381B	0.9819	0.9813	0.9751	0.9809	0.9802	—
	TP: CN104925837B	0.8071	0.8166	0.8027	0.8699	0.7290	3
3	RP: CN105846004A	0.9771	0.9718	0.9759	0.9742	0.9785	—
	TP: CN114940296B	0.8026	0.5463	0.7084	0.7422	0.8657	4
4	RP: CN113745685A	0.9841	0.9735	0.9698	0.9844	0.9667	—
	TP: CN114940296B	0.8255	0.5955	0.7944	0.8211	0.8393	6
5	RP: CN117613439A	0.9794	0.9754	0.9763	0.9771	0.9747	—
	TP: CN106191445B	0.8325	0.7610	0.7976	0.2320	0.7931	5
6	RP: CN107256966B	0.9775	0.9678	0.9767	0.9859	0.9842	—
	TP: CN104810489B	0.2129	0.7867	0.6721	0.8169	0.1733	7
8	RP/TP: CN107331859B	0.9822	0.9745	0.9689	0.9852	0.9836	1
10	RP: CN109390631A	0.9822	0.9777	0.9766	0.9829	0.9825	—
	TP: CN105336953B	0.7520	0.7972	0.7780	0.8315	0.1941	5
11	RP: CN107302075A	0.9797	0.9720	0.9766	0.9816	0.9830	—
	TP 1: CN111697189B	0.8976	0.6399	0.1888	0.7295	0.1786	2
	TP 2: CN111816826B	0.8192	0.5172	0.8067	0.6912	0.8009	9
	TP 3: CN214203906U	0.8924	0.7404	0.7459	0.6553	0.7906	5
	TP 4: CN107195838B	0.8097	0.6162	0.7572	0.7261	0.8962	6
13	RP/TP 1: CN104269555B	0.9851	0.9828	0.9674	0.9859	0.9851	1
	TP 2: CN105261734B	0.9862	0.9836	0.9660	0.9865	0.9855	2
16	RP/TP: CN114256457B	0.9825	0.9719	0.9733	0.9834	0.9730	1
19	RP: CN111477986B	0.9819	0.9769	0.9757	0.9761	0.9763	—
	TP 1: CN108940428B	0.8286	0.7944	0.8084	0.7560	0.9548	4
	TP 2: CN108963376B	0.8648	0.7418	0.8122	0.7725	0.9548	2
21	RP: CN117026315B	0.9795	0.9771	0.9707	0.9686	0.9747	—
	TP: CN106350836B	0.8065	0.7370	0.7298	0.8638	0.8133	10

“—” in the ranking column indicates that the entry is a recommended patent rather than a transferred patent. “RP” denotes Recommended Patent; “TP” denotes Transferred Patent.

presents the recommended patents corresponding to the 12 complete enterprise technology demands along with their supply–demand content conformity scores. For each demand, the element-level conformity scores of the enterprise’s actual transferred patents and their rankings within the algorithm’s recommendation sequence are also reported.

(1) Overall Accuracy Analysis of PR-CTD Recommendation Results

The patents recommended by PR-CTD fully cover all five core element types involved in each enterprise demand, with no instances of pseudo-matching arising from cross-patent element conflation. This demonstrates that the algorithm’s constraint mechanism for patent element completeness operates effectively during the candidate selection stage.

In terms of score distribution, the five element scores of the recommended patents are consistently above 0.96, with an overall mean of approximately 0.98. The lowest individual element score is 0.9667 for the application element of Demand 4, and the highest is 0.9859 for the product element of Demand 6. The concentration of scores indicates that PR-CTD can recommend patents that are highly aligned with enterprise demands at the element level using the patent supply knowledge graph. The five element scores of each recommended patent are well-balanced, with no single element dimension noticeably dragging down the overall score. This suggests that the recommendation results are not inflated by exceptionally high scores on isolated elements, but rather reflect a comprehensive multi-dimensional matching outcome across all five element types. This scoring pattern is consistent with the design of PR-CTD, which imposes holistic constraints and multi-dimensional scoring across all five elements, confirming that the algorithm selects patents based on overall supply–demand element alignment rather than partial matching on a single element.

(2) Score Comparison Between Recommended Patents and Actual Transferred Patents

Comparing the element scores of recommended patents against those of the actual transferred patents under the same demand reveals a substantial gap in supply–demand content conformity. The mean element conformity score of recommended patents is approximately 0.98, while the corresponding scores of actual transferred patents are concentrated in the range of 0.60 to 0.90, representing an overall gap of approximately 0.10 to 0.35. This pattern is consistently observed across all 12 demands, indicating that actual transferred patents generally exhibit lower supply–demand semantic matching than algorithm-recommended patents. The element-level scores display the following characteristics.

First, the method element scores of actual transferred patents are notably low. For example, the method element scores of the actual transferred patents corresponding to Demands 3 and 4 are only 0.5463 and 0.5955, respectively, and the method element scores of all four actual transferred patents for Demand 11 fall below 0.75. By contrast, the method element scores of the recommended patents for all demands exceed 0.97, indicating that the method element is the primary dimension differentiating recommended patents from actual transferred patents.

Second, the application element scores are the most dispersed. Some actual transferred patents receive extremely low application element scores, such as 0.1733 for Demand 6, 0.2320 for Demand 5, and 0.1941 for Demand 10, while others approach or even exceed the level of the recommended patents, such as the two actual transferred patents for Demand 19, which both receive application element scores of 0.9548. Cross-referencing the supply–demand relationship types in Table 3, the analysis finds that the low-scoring samples correspond to industry–academia–research collaborations and entities within the same holding structure. In these cases, the motivation for patent transfer is more likely driven by collaborative agreements or intra-group technology allocation rather than precise matching of application scenarios. Even when there is a mismatch between the patent’s application scope and the enterprise’s stated demand, the transfer can still proceed as long as the core technical direction is satisfactory, after which the enterprise may adapt the application scope to its actual production conditions. This suggests that the application element does not function as a binding constraint in real-world transfer decisions, and that enterprises maintain a degree of flexibility in defining application scenarios.

Third, the score gap for the material and product elements is relatively small. The material and product element scores of most actual transferred patents fall in the range of 0.75 to 0.90, with a smaller gap compared to recommended patents than observed for the method and application elements. This indicates that actual transferred patents are more closely aligned with enterprise demands on the material and product dimensions. This scoring pattern suggests that enterprises prioritize material and product compatibility in their transfer decisions, first assessing whether the technology is applicable in terms of materials and products before considering methodological adjustments based on their own operational conditions.

(3) Patent Ranking Comparison and Analysis of Discrepancies

Regarding recommendation accuracy, the top-ranked patents recommended by PR-CTD for Demands 8, 13, and 16 are identical to the actual transferred patents. The five-element supply–demand conformity scores for these three demands all exceed 0.96, with well-balanced element score distributions and no notable weaknesses. These successful cases demonstrate that when the technical directions of the supply and demand sides are highly aligned, the demand expression is complete, and the supply–demand relationship has a solid technological foundation, the recommendations produced by PR-CTD are consistent with the enterprise’s actual transfer decisions, reflecting strong algorithmic accuracy in such scenarios.

Among the cases where the top recommendation does not match the actual transferred patent, all actual transferred patents rank within the top 10 results, indicating that PR-CTD successfully places actual transferred patents within the high-relevance result range and that the recommended patents retain practical reference value overall. The fact that actual transferred patents do not always rank first reflects that enterprise transfer decisions are not determined solely by supply–demand semantic matching, and that other factors beyond matching quality may influence the outcome. Drawing on the supply–demand relationship types in Table 3 and the element-level scores of each sample, the non-matching cases are categorized into the following three groups, each analyzed to identify the reasons for the ranking discrepancy between recommended and actual transferred patents.

The first group comprises samples with capital relationships (Demand IDs 2, 6, and 21). For Demand 2, the demand side is a wholly owned subsidiary of the supply side, and the actual transferred patent ranks third, with element scores in the range of 0.73 to 0.87, approximately 0.10 to 0.20 below the recommended patent. The relatively uniform gap across all five elements indicates that the transferred patent exhibits meaningful semantic alignment with the demand, and the lower ranking is more likely attributable to intra-group patent allocation than to insufficient supply–demand content matching. For Demand 6, the supply and demand parties belong to the same technology group, and the actual transferred patent ranks seventh. Its material element score (0.2129) and application element score (0.1733) are extremely low, reflecting a mismatch between the patent’s technical focus (a shaping device for coin-type battery casings) and the enterprise’s demand (cathode process optimization for a TPMS lithium micro-power source) on these two element dimensions. The transfer is primarily driven by intra-group asset allocation rather than supply–demand content alignment. For Demand 21, the supply and demand parties belong to the same holding group, and the actual transferred patent ranks tenth, with element scores concentrated in the range of 0.73 to 0.86, reflecting the influence of intra-group technology distribution on the actual transfer decision.

The second group comprises samples with collaborative relationships (Demand IDs 3, 4, 5, and 10). For Demands 3 and 4, the supply and demand parties have previously collaborated in the field of intelligent disassembly of retired lithium batteries, and the actual transferred patents rank fourth and sixth, respectively. The method element scores of the transferred patent are notably low, while the material and application element scores are relatively higher, indicating that the patent aligns with the demand on material and application dimensions but exhibits a certain gap in terms of methodological expression. For Demand 5, the supply and demand parties have a documented industry–academia–research collaboration, and the actual transferred patent ranks fifth, with a product element score of only 0.2320, reflecting a gap between the patent’s product orientation and the enterprise’s product requirements. Nevertheless, the material and method element scores are relatively high, suggesting that the patent’s overall technical direction is aligned with the demand. Demand 10 presents a similar situation, with the actual transferred patent ranking fifth and an application element score of only 0.1941, indicating a mismatch between academic research outputs and enterprise application requirements, as well as certain challenges in application scenario adaptation during patent commercialization.

The third group comprises samples with no identifiable prior association (Demand IDs 11 and 19). For Demand 19, the two actual transferred patents rank second and fourth, respectively, the highest rankings among all non-matching samples. Their element scores are relatively balanced, with an overall range of 0.74 to 0.96, representing the smallest gap relative to the recommended patents. This suggests that in the absence of organizational ties, enterprise transfer decisions are more closely guided by the intrinsic supply–demand content alignment, resulting in the highest degree of consistency with the algorithmic logic. For Demand 11, four actual transferred patents are involved, ranking second, fifth, sixth, and ninth, respectively. While certain patents receive low efficacy (0.1888) and application (0.1786) element scores, all four actual transferred patents fall within the top 10 recommendations, indicating that the algorithm demonstrates strong capability in identifying the patents that enterprises actually transfer.

In summary, at the quantitative level, the patents recommended by PR-CTD consistently outperform actual transferred patents in five-element conformity scores, and the recall rate of actual transferred patents within the recommended candidate sets is high. The cases where PR-CTD does not rank the actual transferred patent first can be attributed to multiple factors. In addition to supply–demand matching quality, enterprise transfer decisions are also influenced by factors such as collaborative relationships, intra-group technology allocation, and technology maturity. The impact of these non-matching factors is particularly pronounced in samples where the supply and demand parties have capital relationships or prior technical collaboration.

4.5. Empirical Evaluation of the Fuzzy Recommendation Algorithm for Incomplete Demands

For incomplete enterprise technology demands, this study employs the FR-ITD algorithm to conduct an empirical analysis of patent recommendation. Incomplete demands are characterized by the absence of one or more elements such as material, method, or efficacy, making it difficult to accurately recommend relevant patents based solely on element-level supply–demand matching. During the matching process, FR-ITD supplements the computation of supply–demand content conformity for identified elements with a demand category fitness score, which serves as an additional evaluation metric to constrain the technical direction of recommended patents.

Table 9. Supply–demand content matching degree of FR-ITD recommended patents and ranking of actually commercialized patents.

Demand ID	Patent	Supply–Demand Content Conformity					Demand Category Fitness Score	Actual Transferred Patent Ranking
		Material	Method	Efficacy	Product	Application
1	RP 1: CN112688029B	—	0.8913	0.8197	0.8923	0.8962	0.9460	—
	RP 2: CN112688029A	—	0.8913	0.8197	0.8923	0.8962	0.9460
	RP 3: CN103342842B	—	0.7977	0.8812	0.8771	0.8962	0.9406
	RP 4: CN113594632A	—	0.8609	0.8801	0.7426	0.8882	0.9503
	RP 5: CN112004592A	—	0.8563	0.8861	0.7676	0.8829	0.9449
	TP: CN103413907B	—	0.7762	0.7636	0.8398	0.1786	0.9078	7
7	RP 1: CN120199807B	0.7516	—	0.8531	0.9706	0.9024	0.9131	—
	RP 2: CN114975940A	0.8034	—	0.8049	0.9376	0.9134	0.8972
	RP 3: CN215299358U	0.8021	—	0.7324	0.9069	0.9069	0.9136
	RP 4: CN113130862A	0.8713	—	0.6659	0.8924	0.9222	0.8967
	RP 5: CN107068990A	0.7981	—	0.7762	0.9706	0.8898	0.8857
	TP: CN209001087U	0.8927	—	0.7797	0.9003	0.1981	0.8711	6
9	RP 1: CN117334878A	0.9006	—	0.8667	0.9755	0.8989	0.9331	—
	RP 2: CN119118135B	0.8547	—	0.8409	0.9214	0.8989	0.9421
	RP 3: CN119118135A	0.8547	—	0.8409	0.9214	0.8989	0.9421
	RP 4/TP: CN116083927B	0.8842	—	0.8543	0.9560	0.8989	0.9284	4
	RP 5: CN109428064A	0.8630	—	0.8574	0.9147	0.8989	0.9366	—
12	RP 1: CN110395032A	—	0.8079	0.8500	0.9142	0.8462	0.9278
	RP 2: CN215432262U	—	0.8369	0.8558	0.8327	0.9173	0.9006
	RP 3/TP 1: CN214203906U	—	0.8079	0.7487	0.8876	0.8934	0.9108	3
	RP 4: CN209735962U	—	0.7992	0.8137	0.7777	0.8824	0.9174	—
	RP 5: CN202289679U	—	0.8461	0.8713	0.7422	0.9240	0.8961
	TP 2: CN111697189B	—	0.6812	0.1706	0.8906	0.1743	0.9219	6
14	RP 1: CN109975381A	—	0.8891	0.8596	0.9066	0.8699	0.8889	—
	RP 2: CN106532148A	—	0.8168	0.8193	0.9158	0.8411	0.9087
	RP 3: CN112382785A	—	0.8213	0.8283	0.9039	0.8747	0.8999
	RP 4: CN113176509A	—	0.7912	0.8909	0.9039	0.8520	0.8933
	RP 5/TP: CN109301103B	—	0.8234	0.8172	0.8959	0.8259	0.9045	5
15	RP 1: CN118231755B	0.9012	0.8060	—	0.9521	0.8598	0.8870	—
	RP 2: CN108288737A	0.8465	0.7154	—	0.9484	0.8142	0.8989
	RP 3: CN108288737B	0.8465	0.7154	—	0.9484	0.8142	0.8989
	RP 4: CN102009972B	0.9374	0.7502	—	0.8450	0.8598	0.8890
	RP 5: CN1884046A	0.9143	0.7828	—	0.8320	0.8598	0.8800
	TP 1: CN113353958B	0.9202	0.7157	—	0.9009	0.8986	0.7930	8
	TP 2: CN114956021B	0.9203	0.7247	—	0.9165	0.1972	0.8012	10
17	RP 1: CN113964293B	0.8364	—	0.8123	0.8823	0.9078	0.9505	—
	RP 2: CN113964293A	0.8364	—	0.8123	0.8823	0.9078	0.9505
	RP 3: CN114976266A	0.8491	—	0.8180	0.8740	0.9078	0.9364
	RP 4/TP: CN112599782A	0.8065	—	0.8515	0.8088	0.9147	0.9322	4
	RP 5: CN115800438A	0.8160	—	0.8757	0.8166	0.9078	0.9172	—
18	RP 1: CN209169329U	0.8947	—	—	0.8211	—	0.9347	—
	RP 2: CN104037386A	0.8299	—	—	0.8797	—	0.9298
	RP 3: CN106233525B	0.8804	—	—	0.8860	—	0.9108
	RP 4: CN106233525A	0.8804	—	—	0.8860	—	0.9108
	RP 5: CN206758525U	0.8717	—	—	0.8221	—	0.9335
	TP: CN218677247U	0.8180	—	—	0.9083	—	0.8893	11
20	RP 1/TP 1: CN113956282B	0.8674	—	0.8661	0.8639	0.9278	0.9234	1
	RP 2: CN116272844B	0.7752	—	0.8092	0.8521	0.9328	0.9206	—
	RP 3/TP 2: CN113745660B	0.8400	—	0.8155	0.8984	0.9075	0.9112	3
	RP 4: CN115650261B	0.8263	—	0.8475	0.8899	0.9170	0.9103	—
	RP 5: CN106829908B	0.7819	—	0.7871	0.8639	0.9328	0.9379	—
	TP 3: CN119695238B	0.7784	—	0.7153	0.8570	0.8315	0.8234	6
	TP 4: CN119674234B	0.7458	—	0.7548	0.8760	0.8134	0.8307	10
22	RP 1: CN117559079A	0.7907	—	0.8090	0.8975	0.9544	0.9053	—
	RP 2: CN116454530B	0.8109	—	0.6686	0.9186	0.9392	0.9121
	RP 3: CN107742688B	0.7719	—	0.7264	0.9575	0.9063	0.9049
	RP 4: CN107221625A	0.8009	—	0.6768	0.9086	0.9392	0.9090
	RP 5: CN104733676A	0.7558	—	0.7596	0.9293	0.9392	0.8982
	TP 1: CN102290549B	0.8485	—	0.6244	0.9602	0.8921	0.8595	9
	TP 2: CN102267229B	0.8666	—	0.6589	0.9530	0.9425	0.8450	13

“—” indicates that the corresponding element type is absent from the demand or is not applicable to the patent entry. “RP” denotes Recommended Patent; “TP” denotes Transferred Patent.

presents the supply–demand content conformity scores, demand category fitness scores, and the rankings of enterprises’ actual transferred patents within the recommendation results for incomplete demands. The recommendation performance of FR-ITD is analyzed from three perspectives: overall recommendation outcomes, content analysis of matched cases, and content analysis of unmatched cases.

(1) Overall Accuracy Analysis of FR-ITD Recommendation Results

In terms of recommendation accuracy, the 10 incomplete demands collectively involve 14 actual transferred patents. FR-ITD successfully retrieves 6 transferred patents within the Top-5 recommendation results, distributed across Demand 9 (1 patent), Demand 12 (1 patent), Demand 14 (1 patent), Demand 17 (1 patent), and Demand 20 (2 patents). The remaining 8 transferred patents rank between 6th and 13th, indicating that the algorithm places actual transferred patents within a high-relevance result range, thereby retaining substantial practical reference value.

In terms of overall score distribution shown in Table 9, the supply–demand content conformity scores of FR-ITD recommended patents for identified elements are consistently high, with the majority concentrated in the range of 0.80 to 0.97. This demonstrates that the algorithm can accurately identify demand content and recommend matching patents even under conditions of incomplete element information. Regarding demand category fitness, the scores of recommended patents are generally above 0.90, with some approaching 0.95, indicating that the demand category identification mechanism effectively constrains the technical direction of candidate patents and ensures consistency between recommendation results and enterprise demands.

Compared with PR-CTD, which targets complete demands, FR-ITD produces a Top-5 candidate set. Because incomplete demands have ambiguous boundaries due to missing elements, a single optimal patent cannot comprehensively cover all potential technical directions. Recommending a set of five candidates provides enterprises with a diverse range of options, striking a balance between demand coverage and recommendation precision.

(2) Score Analysis of Successfully Matched Patents

Among the demands for which recommendations are successful, the final ranking of each patent is jointly determined by its supply–demand content conformity score and its demand category fitness score.

For Demand 20, the enterprise’s actual transferred patents CN113956282B and CN113745660B rank first and third, respectively. Demand 20 concerns high-voltage electrolyte technology, with the identified elements being material, efficacy, and application, while the method and product elements are absent. CN113956282B achieves balanced scores across four identifiable element types: material (0.8674), efficacy (0.8661), product (0.8639), and application (0.9278), with a demand category fitness score of 0.9234, covering the technical directions of electrolyte preparation and lithium salt purification. CN113745660B receives relatively high material (0.8400) and product (0.8984) element scores, with a category fitness score of 0.9112.

For Demand 17, the matched patent CN112599782A (ranked 4th) achieves the highest application element score (0.9147) among all candidate patents for this demand, as well as the highest demand category fitness score (0.9322). This indicates that the patent is highly aligned with the enterprise’s demand in the 4.45 V fast-charging lithium-ion battery application scenario, and that the demand category mechanism successfully identifies the underlying high-voltage fast-charging requirement, compensating for the information deficit caused by the absence of the method element. For Demand 9, the matched patent CN116083927B achieves a product element score of 0.9560, the highest among all candidate patents for this demand, along with a demand category fitness score of 0.9284. These results reflect that the algorithm accurately identifies the silicon-carbon anode material technology direction and recommends patents oriented toward high-energy-density lithium storage performance.

Demands 12 and 14 both lack the material element. The matched patents CN214203906U and CN109301103B each receive efficacy and product element scores above 0.87, with category fitness scores of 0.9108 and 0.9045, respectively. The technical scope of Demand 12 focuses on polyolefin microporous base membranes and pore structure control of separators. CN214203906U is a polyolefin microporous membrane with a specific pore structure, whose pore structure regulation direction directly matches the enterprise’s demand. Demand 14 incorporates supplementary technical categories including thermal runaway protection and precision heat dissipation channel design, and CN109301103B aligns with the demand category in terms of precision structural component design for prismatic power batteries. The common feature of these two matched patents is that the demand category supplementation compensates for the missing elements, enabling the algorithm to make recommendations.

(3) Score Gap Analysis for Unmatched Patents technically appropriate scope.

The 8 unmatched transferred patents rank between 6th and 13th in the algorithm results, indicating that FR-ITD identifies a supply–demand matching relationship between these patents and the enterprise demands, but their composite scores are relatively lower and they do not enter the Top-5 recommendation results. This demonstrates that the algorithm’s recommendation direction is not misaligned; the unmatched patents remain within the candidate range and retain certain reference value for enterprise decision-making. The reasons for the ranking shortfall of unmatched transferred patents are analyzed below in conjunction with supply–demand relationships, content conformity scores, and demand category fitness scores.

First, some transferred patents exhibit low application element matching. The application element score of the transferred patent CN103413907B for Demand 1 is only 0.1786, while the application element scores of the recommended patents for the same demand are concentrated in the range of 0.87 to 0.90. Similarly, the application element score of CN209001087U for Demand 7 is 0.1981, compared with recommended patent scores of 0.89 to 0.92. The transferred patent CN114956021B for Demand 15 receives an application element score of 0.1972, and the transferred patent CN111697189B for Demand 12 scores poorly on both the efficacy element (0.1706) and the application element (0.1743). In terms of supply–demand relationships, the demand side of Demand 1 is a wholly owned subsidiary of the supply side, as is the demand side of Demand 15, where the organizational relationship between the parties determines the transfer decision. For Demand 12, although the supply and demand parties have no capital relationship, the transferred patent CN111697189B is oriented toward polyolefin microporous base membranes and is closely related to the demand side’s products. This pattern is consistent with the finding from the PR-CTD empirical evaluation that enterprises tend to apply relatively flexible matching requirements to the application element in transfer decisions.

Second, low demand category fitness scores negatively affect the ranking of transferred patents. The category fitness scores of the two transferred patents for Demand 15 are only 0.7930 and 0.8012, both lower than the recommended patents for the same demand. The category fitness scores of the two transferred patents for Demand 22 are 0.8595 and 0.8450, below the recommended patents’ scores of 0.90 to 0.91. For Demand 20, transferred patents 3 and 4 have category fitness scores of 0.8234 and 0.8307, respectively, both lower than the matched transferred patents 1 and 2. Considering the supply–demand relationships, the supply and demand parties for Demands 15, 20, and 22 are all connected through wholly owned holding or subsidiary relationships, suggesting that the transfer behavior may be influenced by intra-group technology allocation. Under such circumstances, demand category fitness is not necessarily a primary consideration, and patents with lower category fitness scores can still be transferred.

Third, non-content matching factors may influence patent transfer decisions. For Demand 18, only two element types—material and product—are available, leaving the algorithm with limited semantic information for matching. The demand category fitness score of the transferred patent CN218677247U (0.8893) is lower than that of the recommended patents, and the supply and demand parties have no identifiable association, with the transfer taking the form of an individual patent assignment. It is therefore plausible that the enterprise considered factors such as technology maturity or specific technical requirements in its transfer decision. For Demand 22, the efficacy element scores of the two transferred patents are low; however, given that the supply and demand parties are in a wholly owned subsidiary relationship and the transfer represents a patent reversion from a subsidiary to its parent company, category fitness and efficacy indicators are unlikely to be the primary decision criteria. The deviation between the recommendation results and the actual transfers in this case is, to some extent, beyond the controllable scope of the algorithm.

In summary, FR-ITD demonstrates strong recommendation performance in the empirical evaluation of 10 incomplete demands. Among the 14 actual transferred patents, 6 enter the Top-5 results and the remaining 8 all rank within the top 13, demonstrating that the algorithm can identify demand content and place actual transferred patents within a high-relevance candidate range even under conditions of incomplete element information. The common characteristics of matched patents are high supply–demand content conformity scores for identified elements and high demand category fitness scores, confirming that the demand category mechanism plays a critical role in compensating for missing elements and constraining recommendation direction. Unmatched transferred patents may be affected by low application element matching, insufficient demand category fitness, or non-content matching factors that fall outside the algorithm’s controllable scope. Compared with PR-CTD, FR-ITD targets incomplete demands with greater information ambiguity, and adopts a Top-5 candidate set in place of a single optimal recommendation, striking a balance between demand coverage and recommendation precision.

(4) Diversity Analysis of Recommendation Results

To evaluate whether FR-ITD fulfills its design objective of generating diversified candidate sets, this study examines the supply solution types associated with each recommended patent. Patent supply solution types (S1–S15) are derived from a TRIZ-based classification framework developed in a closely related study [46], with full definitions provided in Appendix A. Each recommended patent is mapped to a combination of solution types, and recommendation diversity is defined as the breadth of distinct solution types covered by the Top-5 candidate set for a given demand, reflecting the range of technically differentiated implementation pathways offered to the enterprise.

Table 10. Comparison of Patent Supply Solution Types in FR-ITD Recommendation Results.

Demand ID	Patent	Patent Supply Solution Types
1	RP 1: CN112688029B	S1; S9; S10
	RP 2: CN112688029A	S1; S3; S7
	RP 3: CN103342842B	S1; S2; S5
	RP 4: CN113594632A	S1; S5; S7
	RP 5: CN112004592A	S1; S3; S9
	TP: CN103413907B	S1; S3; S6
7	RP 1: CN120199807B	S1; S5; S6
	RP 2: CN114975940A	S1; S5; S6
	RP 3: CN215299358U	S1; S4; S11
	RP 4: CN113130862A	S1; S2; S5
	RP 5: CN107068990A	S1; S5; S6
	TP: CN209001087U	S8; S9; S7
9	RP 1: CN117334878A	S1; S5; S6
	RP 2: CN119118135B	S1; S5; S6
	RP 3: CN119118135A	S1; S5; S6
	RP 4/TP: CN116083927B	S1; S5; S6
	RP 5: CN109428064A	S1; S5; S6
12	RP 1: CN110395032A	S1; S5; S7
	RP 2: CN215432262U	S3; S10; S14
	RP 3/TP 1: CN214203906U	S1; S3; S7
	RP 4: CN209735962U	S7; S15
	RP 5: CN202289679U	S3; S5; S7
	TP 2: CN111697189B	S1; S3; S7
14	RP 1: CN109975381A	S5; S6; S7
	RP 2: CN106532148A	S10; S7; S15
	RP 3: CN112382785A	S8; S10; S7
	RP 4: CN113176509A	S3; S12; S10
	RP 5/TP: CN109301103B	S3; S8; S12
15	RP 1: CN118231755B	S1; S5; S7
	RP 2: CN108288737A	S15; S7; S5
	RP 3: CN108288737B	S15; S7; S5
	RP 4: CN102009972B	S1; S5; S7
	RP 5: CN1884046A	S1; S5; S6
	TP 1: CN113353958B	S3; S5; S7
	TP 2: CN114956021B	S2; S5; S15
17	RP 1: CN113964293B	S1; S6; S7
	RP 2: CN113964293A	S1; S6; S7
	RP 3: CN114976266A	S1; S5; S6
	RP 4/TP: CN112599782A	S1; S6; S9
	RP 5: CN115800438A	S4; S6; S10
18	RP 1: CN209169329U	S4; S7; S10
	RP 2: CN104037386A	S1; S3; S8
	RP 3: CN106233525B	S1; S6; S7
	RP 4: CN106233525A	S1; S6; S7
	RP 5: CN206758525U	S8; S9; S10
	TP: CN218677247U	S8; S10; S9
20	RP 1/TP 1: CN113956282B	S1; S2
	RP 2: CN116272844B	S1; S2; S15
	RP 3/TP 2: CN113745660B	S1; S2
	RP 4: CN115650261B	S1; S5; S6
	RP 5: CN106829908B	S1; S5; S7
	TP 3: CN119695238B	S1; S2; S6
	TP 4: CN119674234B	S1; S2; S7
22	RP 1: CN117559079A	S1; S5; S7
	RP 2: CN116454530B	S1; S5; S7
	RP 3: CN107742688B	S1; S9; S10
	RP 4: CN107221625A	S1; S5; S7
	RP 5: CN104733676A	S1; S5; S6
	TP 1: CN102290549B	S1; S5; S6
	TP 2: CN102267229B	S1; S5; S7

“RP” denotes Recommended Patent; “TP” denotes Transferred Patent.

presents the recommended patents and their corresponding solution type combinations for the FR-ITD cases.

The solution type distributions of FR-ITD recommendation results provide a structured basis for assessing recommendation diversity. By mapping each recommended patent to its corresponding supply solution types, it becomes possible to evaluate not only whether recommended patents are technically relevant but also whether they collectively represent a range of distinct implementation pathways. For most demands, the Top-5 candidate patents span three to five distinct solution types, covering complementary technical dimensions rather than converging on a single approach. This breadth reflects the design intent of FR-ITD, which expands sparse demand information through multi-source enterprise data and augmented query contexts to generate candidate sets that represent a range of technically viable pathways rather than replicating the same solution.

The diversity patterns observed across cases are consistent with the degree of element completeness in each demand. Demand 1 illustrates a case of strong within-set diversity, with Top-5 results spanning structural innovation (S9, S10), process optimization (S3), and material modification (S2, S5), collectively offering the enterprise options that differ not only in technical content but also in implementation pathway. Demand 12 and Demand 14 exhibit similarly broad coverage, with recommended patents distributed across material compounding (S1), process reconfiguration (S3), intelligent control (S10), safety management (S12), and lifecycle design (S15), reflecting the wide candidate space that FR-ITD generates when demand elements are limited and the solution boundary is loosely defined. Demand 18 further demonstrates cross-dimension diversity, where Top-5 results include patents oriented toward intelligent thermal management (S4), production process optimization (S3), electrochemical methods (S6), and structural safety design (S8, S10), providing the enterprise with options spanning both process-level and product-level solutions.

By contrast, Demand 9 exhibits lower within-set diversity, with Top-5 results concentrated around S1, S5, and S6. This outcome is not a limitation of the method but rather a reflection of the demand itself: the available elements already point toward a relatively specific technical direction, and FR-ITD appropriately narrows the candidate space in response. This adaptive behavior, producing broader coverage for informationally sparse demands and more focused results for directionally constrained ones, distinguishes FR-ITD from methods that apply uniform retrieval logic regardless of demand characteristics.

It is also noteworthy that in several cases the actual transferred patents correspond to solution types that fall within the range covered by the Top-5 recommendations, even when an exact match is not achieved. For example, in Demand 1 the transferred patent (S1, S3, S6) shares solution types with multiple recommended patents, and in Demand 18 the transferred patent (S8, S10, S9) aligns closely with the solution type profile of the fifth recommended patent. This overlap suggests that FR-ITD successfully captures the relevant solution space even under incomplete demand conditions, and that the actual transfer decision can be understood as a selection from within the diversified candidate set rather than a departure from it.

The diversity analysis further confirms that FR-ITD generates candidate sets spanning multiple technically distinct solution pathways, providing enterprises with meaningful comparative options that a single-output or purely similarity-ranked approach would not afford.

5. Discussion and Conclusions

5.1. Discussion

This study developed a differentiated patent recommendation framework to address heterogeneous enterprise technological demands and empirically validated the proposed methods in the lithium battery technology domain. The results show that the two algorithms, PR-CTD and FR-ITD, effectively identify patents that match enterprise technological demands under different information conditions. The recommendation results provide meaningful references for patent transfer decisions in lithium battery enterprises.

For complete technological demands, the PR-CTD algorithm evaluates supply–demand content conformity across five technological elements: material, method, efficacy, product, and application. The empirical results show that the average five-element conformity score of the recommended patents was approximately 0.98, with most scores above 0.96. The balanced distribution across the five elements indicates that the algorithm successfully constrains the matching process at multiple technological dimensions and avoids bias caused by a single dominant element. In several cases, the top-ranked recommended patent was identical to the patent that was actually transferred by the enterprise. In the remaining cases, the transferred patents still appeared within the top-ranked candidate range, indicating that the algorithm was able to place the real transferred patents within a highly relevant solution space.

Analysis of element-level scores reveals clear industry-specific patterns. The gaps in matching scores were smallest for the material and product elements, while the largest differences appeared in the method element. This pattern reflects technological characteristics of the lithium battery industry. Enterprises typically prioritize compatibility with core material systems and product structures when acquiring external technologies. Process methods are often adapted to existing production conditions after the transfer rather than directly adopted from the original patent. The application element also shows relatively dispersed scores, which reflects the diversity of application scenarios across the lithium battery supply chain and the flexibility with which enterprises adjust application contexts after technology transfer.

For incomplete technological demands, the FR-ITD algorithm addresses the lack of element information through demand category supplementation and semantic expansion. The empirical results show that the recommended patents achieved high conformity scores for the identified demand elements, while demand category fitness effectively constrained the technological direction of the recommended patents. Several actual transferred patents appeared within the Top-5 recommendation results, and the remaining cases were still located within the high-relevance candidate range. These results indicate that the algorithm can still identify relevant technological solutions even when the original demand description contains incomplete information.

Incomplete demand expression is common in the lithium battery industry. Enterprises frequently describe desired performance outcomes while leaving implementation methods unspecified. The FR-ITD strategy recommends a Top-5 candidate set rather than a single patent. This approach allows enterprises to compare alternative technological pathways and is therefore better suited to situations in which demand boundaries are uncertain.

The differences between recommendation results and actual transfer decisions also reveal several characteristics of technology transfer behavior in the lithium battery industry. When no organizational relationships exist between the supply and demand parties, technology content matching becomes the dominant factor in transfer decisions. In these cases, the recommendation results show high consistency with actual transfers. However, when capital ties or industry–academia collaboration relationships are present, patent transfer decisions may be influenced by strategic considerations, organizational structures, or cooperation mechanisms. These factors may lead to differences between algorithmic recommendations and real transfer outcomes.

In the context of global energy transition and low-carbon development, lithium battery technology plays a critical role in supporting electric vehicles and large-scale energy storage systems. Efficient identification of relevant patents from large patent databases is therefore essential for accelerating technological diffusion and innovation in this sector. The recommendation framework proposed in this study provides a structured tool that helps enterprises identify external technological resources from large patent corpora. This capability can reduce search costs, support technology transfer decisions, and promote sustainable technological innovation within the lithium battery industry.

5.2. Conclusions

This study proposes a patent recommendation framework driven by enterprise technological demands to address supply–demand matching challenges in technology-intensive industries. The framework integrates three key components: enterprise technological demand identification, patent supply knowledge graph construction, and differentiated recommendation algorithms. The effectiveness of the proposed framework was empirically validated in the lithium battery technology domain.

At the demand identification stage, enterprise technological demands are decomposed into five elements: material, method, efficacy, product, and application. Pre-trained language models and named entity recognition techniques were used to automatically extract these elements from demand texts, enabling a structured representation of enterprise technological needs. Based on the completeness of extracted elements, demands were classified into complete and incomplete types. For incomplete demands, additional information from enterprise industry background, historical technology development, and technological capability requirements was integrated to supplement demand categories.

At the patent supply knowledge graph construction stage, element entities were extracted from patent texts through a combined approach that integrates structured field extraction and large language model recognition. Knowledge graph embedding methods were then used to represent the semantic relationships among entities, forming a patent supply knowledge graph capable of supporting multi-dimensional supply–demand matching.

At the recommendation stage, two algorithms were developed for different demand conditions. The PR-CTD algorithm targets complete technological demands and evaluates supply–demand content conformity across the five demand elements. The FR-ITD algorithm targets incomplete technological demands and combines element conformity with demand category fitness. It employs retrieval-augmented generation techniques to dynamically incorporate category information and outputs a Top-5 candidate set to balance recommendation coverage and precision under uncertain demand conditions.

The empirical results demonstrate that the PR-CTD algorithm achieves high element-level matching accuracy for complete demands, while the FR-ITD algorithm effectively identifies relevant patents when demand information is incomplete. The results also reveal several industry-level characteristics of patent transfer decisions in the lithium battery sector. Enterprises place greater emphasis on matching material systems and product structures, while application scenarios are often adjusted after technology acquisition. Organizational relationships and industry–academia collaboration can also influence transfer decisions.

Overall, the proposed framework provides systematic methodological support for enterprises seeking external technological resources from large patent databases. For the lithium battery industry, which is a key enabling technology for electric vehicles and energy storage systems, improving the efficiency of patent transfer and technology diffusion is essential for accelerating technological innovation. The methods proposed in this study contribute to reducing the search cost of identifying suitable patents and support the sustainable development of the lithium battery industry by facilitating more efficient technology transfer.

5.3. Limitations and Future Directions

Although the empirical results demonstrate the effectiveness of the PR-CTD and FR-ITD algorithms under different demand information conditions, several limitations remain and should be considered in future research.

First, the scale and scope of the validation dataset are relatively limited. The demand-driven patent transfer samples used in this study were collected from several provincial technology transfer platforms, and the number of validated cases remains relatively small. Most samples are concentrated in subsectors of the lithium battery supply chain such as cathode materials, anode materials, separators, and electrolytes, while other segments such as battery management systems and disassembly equipment are less represented. Therefore, the generalizability of the results to other technological domains requires further verification. Future research can expand the validation dataset to additional technology-intensive sectors such as advanced materials, intelligent manufacturing, biomedicine, and information technology to evaluate the cross-domain applicability of the proposed recommendation methods.

Second, discrepancies were observed between some recommendation results and the patents that were actually transferred by enterprises. Case analysis indicates that these differences are mainly related to several external factors. In samples involving parent-subsidiary relationships or entities within the same holding structure, patent transfer decisions were driven primarily by internal technology resource allocation rather than pure supply–demand content matching. In samples involving universities or research institutes, research outputs sometimes aligned with enterprise needs at the material or methodological level but showed lower correspondence at the product or application level. This phenomenon reflects the well-known gap between academic research outputs and industrial implementation requirements. Enterprises also tend to apply relatively flexible criteria to application element matching and may adjust application scenarios after acquiring patents. Because the algorithm evaluates the five demand elements with equal weights, some transferred patents with strong alignment in other elements but lower application scores may not be ranked at the top of the recommendation results. In addition, technology maturity, patent pricing, implementation difficulty, and negotiation willingness may influence transfer decisions and fall outside the scope of the current recommendation framework.

Third, the current framework focuses primarily on supply–demand content matching and does not yet incorporate organizational relationships, patent value indicators, or technology maturity factors into the recommendation model. In the lithium battery industry, which is both technology intensive and capital intensive, capital ties and industry–academia collaboration networks often influence technology diffusion and patent commercialization. Future research may integrate organizational relationship networks, patent value metrics, and technology maturity indicators into the recommendation framework to better capture real decision-making processes in patent transfer.

Fourth, the present study assigns equal weights to the five demand elements when calculating supply–demand matching scores. However, empirical results suggest that lithium battery enterprises place greater emphasis on matching material systems and product structures than on application scenarios. This finding indicates that the equal-weight evaluation scheme may not fully reflect actual industrial decision preferences. Future research could develop domain-adaptive weighting mechanisms based on expert knowledge or historical patent transfer data.

Finally, the current framework focuses mainly on algorithmic recommendation and does not yet incorporate interactive mechanisms between the recommendation system and enterprise users. Future studies could introduce user feedback into the recommendation process by recording enterprise responses to recommended patents and dynamically adjusting model parameters based on observed preferences. Improving the interpretability of recommendation results would also help enterprises better understand the technological rationale behind the recommended patents and improve the usability of the system in real technology transfer scenarios.

In the context of global energy transition and low-carbon development, the lithium battery industry plays a critical role in enabling sustainable technologies such as electric vehicles and large-scale energy storage systems. Improving the efficiency of patent transfer and technology diffusion is therefore essential for accelerating green innovation. By expanding validation datasets, incorporating industrial decision factors, and improving interactive recommendation mechanisms, future research can contribute to building more intelligent and sustainable patent recommendation systems that support technological innovation and sustainable development in emerging energy industries.