Research on the Configurational Paths of Collaborative Performance in the Innovation Ecosystem from the Perspective of Complex Systems

Abstract

This study integrates complex systems theory and innovation ecosystem theory to develop a unified framework encompassing the innovation environment, innovation actors, and innovation networks. Using fsQCA and NCA, it examines the impact of cross-layer interactions and the coupling of multiple factors on collaborative performance. Empirical analysis in the field of natural language processing (NLP) demonstrates that no single factor is sufficient to serve as a necessary condition for achieving high-innovation collaboration performance. Innovation actors, as endogenous evolutionary drivers, play a central and catalytic role in the collaboration process. Moreover, under specific conditions, the relationship between the innovation environment and innovation networks exhibits a substitutive effect, with certain capabilities enabling this dynamic. This study extends the theoretical understanding of collaboration pathways within innovation ecosystems and offers practical recommendations for fostering innovation cooperation across different industries and organizations. It achieves this by constructing a “situational type–configuration path” matrix, decision tree, and innovation collaboration performance realization model.

1. Introduction

As scientific research advances alongside developments in artificial intelligence (AI) and other fields, innovation activities are becoming more complex and specialized. This trend makes it increasingly difficult for individual entities, limited by their internal knowledge and resources, to conduct deep innovation independently [1]. As a result, research on collaborative innovation within interdependent, multi-entity innovation ecosystems has attracted increasing attention. From the perspective of innovation ecosystems, institutional collaboration across different layers has shifted from linear-mechanistic innovation models to organic innovation ecosystem models [2]. The advancement of innovation research has enabled innovation ecosystems to achieve value co-creation through multi-entity coordination. By aligning participants around shared value propositions, innovation ecosystems have emerged as a powerful driver of innovation, enabling a structural shift in innovation paradigms [3].

Academic research has systematically examined the factors influencing innovation collaboration from multiple perspectives. Existing studies have primarily focused on the effects of individual dimensions—such as technological similarity, network structure [4], and knowledge diversity [5]—on innovation performance. However, these studies often overlook the synergistic effects of multiple factors within complex innovation ecosystems, limiting our understanding of the underlying mechanisms driving innovation performance. Furthermore, although scholars have examined the influence of multidimensional proximity—including geographical, institutional, and organizational proximity—on innovation collaboration [6,7], they still rely on traditional methods. As a complex system, the innovation ecosystem challenges the assumption that the overall effect equals the sum of individual factor impacts [8]. Characterized by interaction and coordination, the open innovation ecosystem provides a platform for collaborative innovation. Configuration theory, the methodological foundation of fsQCA, aligns conceptually with the complex systems perspective of innovation ecosystems. Configuration analysis enables the examination of multiple levels of key factors—such as the external environment and organizational management—and their combined effects on outcomes [9]. Rooted in configurational thinking, fsQCA critiques the single-factor “net effect” approach [10], and instead explores causal asymmetry, multiple equifinal paths, and conjunctural causation [11]. Therefore, the multi-factor interaction inherent in innovation ecosystems is logically consistent with the “multiple conjunctural causation” emphasized by fsQCA methodology.

This study adopts perspectives from complex systems theory and innovation ecosystem theory to construct a collaborative innovation ecosystem. By integrating fsQCA and NCA, the research explores the diverse modes of innovation collaboration formed by entities across the scientific, technological, and industrial layers, driven by the coupled interactions of multiple factors within the innovation environment, innovation actors, and innovation networks. It uncovers underlying application strategies, addressing the limitations of traditional linear analyses that rely on a single perspective and overcoming the inadequacies of decomposing the collaborative innovation ecosystem.

The study delves into the factors influencing collaborative performance within innovation ecosystems, identifying key combinations of conditions and their path effects through QCA. It challenges the traditional paradigm of causal inference based on single-variable explanations, emphasizing instead the interactive effects between conditions and employing exploratory configurational identification. By analyzing the linkage and coupling mechanisms of multiple antecedent conditions, this research offers a more systematic theoretical framework and multidimensional empirical support for collaborative innovation studies, providing actionable strategic recommendations and implementation pathways for practical applications.

2. Related Research

2.1. Research on Innovation Ecosystems

In 1993, Moore introduced ecological thinking into economic management by proposing the concept of “business ecosystems,” defined as interconnected communities of organizations [12]. Building on this, Adner articulated the core of innovation ecosystems as distinctive collaborative mechanisms [13]. He emphasized that no single actor can independently carry out the entire innovation process. Instead, diverse entities form a pluralistic organizational system based on functional complementarity and interdependence, thus constituting an innovation ecosystem [14]. Within such systems, multilateral dependencies emerge—a structural interdependence that goes beyond simple bilateral cooperation and fosters innovation-oriented organizations with internal order and collaborative norms [15].

During the evolution of innovation ecosystems, interactions among actors have moved beyond traditional static and linear models, forming a complex, organically evolving network with multiple nested layers [16]. This evolution is driven by both exogenous factors, arising from changes in the external environment, and endogenous forces, generated through continuous interactions among actors within the system. Endogenous forces, rooted in human creativity and the intrinsic pursuit of value maximization, are unique to innovation ecosystems and absent in natural ecological systems [2].

Essentially, an innovation ecosystem is a networked construct [17], where innovation activities are distributed across complementary networks composed of diverse resources [18], and their emergence and evolution unfold in intrinsically networked forms [19]. Classical models of innovation ecosystems typically identify three core elements: the innovation environment, innovation actors, and a relational matrix [20]. The term “matrix” is metaphorical, denoting a complex interactive network formed through formal and informal actor linkages that facilitates resource flows and knowledge spillovers [21]. As research paradigms evolve, the matrix concept has gradually been replaced and operationalized as the more precise notion of an innovation network [3]. Granstrand’s definition of innovation ecosystems emphasizes actor interactions as a central analytical focus [22], further highlighting the foundational role of network structures. According to Liu et al. [3], an innovation ecosystem is a collaborative structure focused on achieving innovative outcomes. Within such a system, actors—supported by the innovation environment and guided by shared objectives—integrate diverse ecosystem resources to establish channels for the exchange of information, knowledge, and technology, thereby forming a “shared-success-oriented innovation network. Network formation relies on the development of intra-actor relationships, reflecting both interactive processes and resource flows. Innovation networks link exogenous forces, such as macro-level institutions, with endogenous forces, including actors’ technical and knowledge capacities, forming a transmission network system. Within this system, network relationships provide structural embeddedness, shaping the pathways of resource and information flows as well as the underlying mechanisms driving innovation performance.

Therefore, the evolution of innovation ecosystems is not merely the additive effect of exogenous and endogenous forces, but the outcome of their deeply coupled interactions. Existing studies often reduce innovation cooperation in such complex systems to the sum of individual layers—innovation environment, actors, and networks—thereby overlooking the intricate interdependencies among elements and limiting the explanatory power for the synergistic effects generated by multi-level, multi-component configurations.

2.2. Research on the Influencing Factors of Collaboration from the Perspective of Innovation Ecosystems

2.2.1. Research on the Impact of the Innovation Environment on Innovation Collaboration Performance

(1)

Geographical Proximity

Geographical proximity refers to the spatial distance between collaborating institutions, Shaw & Gilly highlighted the important role of geographical proximity in facilitating innovation interactions [23]. Although advances in information technology have partially reduced spatial constraints, organizations still tend to prefer geographically proximate partners when pursuing collaborative innovation [24]. Excessive distance between collaborating institutions can cause information loss or distortion during knowledge transfer [25], hinder the flow of tacit knowledge, and increase the economic cost of interaction [26]. Liu et al. demonstrated that geographical proximity has a significantly positive effect on innovation complexity and the quality of innovation outputs [27]. However, Cao et al., using stepwise regression to examine the link between geographical proximity and open innovation in organizational nodes of innovation networks, identified a U-shaped relationship [28]. Thus, the relationship between geographical proximity and innovation performance is complex and nonlinear, making it difficult for traditional linear models to accurately capture its underlying mechanism.

(2)

Institutional Proximity

Institutional proximity refers to the degree of similarity between collaborating institutions in terms of institutional frameworks and operational paradigms [29]. The activities of collaborating institutions are strongly shaped by the social institutions and cultural environments in which they are embedded [30]. Alignment in policy orientation reduces institutional barriers, facilitates resource integration, and enhances strategic coordination—thereby improving both the depth and breadth of regional collaboration [31]. Shared institutional arrangements act as a “binding agent” for cooperation by fostering trust and increasing the likelihood of successful collaborative innovation [29]. However, excessive institutional proximity may lead to intensified competition among institutions operating in similar contexts [32], which can offset its positive impact on collaborative innovation. Therefore, excessive institutional similarity does not necessarily enhance collaborative innovation performance.

2.2.2. Research on the Impact of Innovation Actors on Innovation Collaboration Performance

(1)

Technological Proximity

The intersection and integration of disciplines have elevated knowledge to a crucial factor in partner selection among collaborating institutions [33]. Technological proximity refers to the similarity between collaborating parties in their foundational technologies at the knowledge-structure level [34], where technological relationships encompass both similarity and complementarity [35]. A shared technological structure offers collaborating institutions a common knowledge base, facilitating more efficient mutual understanding and knowledge absorption [36], thereby reducing communication costs [37]. However, excessive technological homogeneity can result in path dependency among collaborating institutions, which hinders the development of innovative collaboration. Knowledge recombination is considered a key driver of innovation, and technological heterogeneity can positively affect innovation performance to some extent [38]. Furthermore, some scholars have found that the relationship between technological proximity and collaborative innovation is nonlinear [7], often following an inverted U-shaped curve [36,39].

(2)

Collaboration Tendency

Under constraints like time and cost, collaborating institutions tend to prefer partners with whom they have prior collaborative experience, Previous cooperative experiences are transformed into tacit and explicit organizational knowledge, gradually accumulating into core competencies [40]. This knowledge repository enables organizations to build unique competitive advantages. Trust mechanisms established through past collaborations positively influence collaborative performance [41]. They effectively reduce uncertainties in the collaboration process and mitigate potential opportunistic behaviors. However, these collaborative preferences carry a latent risk of knowledge homogenization. The persistence and stability of such partnerships, if they lead to over-reliance on established models, may reduce receptivity to novel innovations and negatively affect innovation performance [42].

2.2.3. Research on the Impact of Innovation Networks on Innovation Collaboration Performance

(1)

Network Relationship Quantity (NRQ)

Network tie quantity reflects a collaborating institution’s capacity to integrate information resources. Degree centrality measures the number of direct connections a collaborating institution maintains within an innovation network, serving as an indicator of the actor’s network power [43], Institutions occupying central positions in innovation networks possess greater control capability. Ferriani et al. demonstrated that collaborating institutions with more direct partnership ties exhibit stronger abilities to access information and resources [44]. Zhang et al. and Zhao et al. further revealed a positive correlation between centrality and innovation performance [45,46]. However, some scholars argue against a simple linear relationship. Yan et al. identified an inverted U-shaped relationship between network centrality and innovation performance [47]. Within innovation ecosystems, the complex nonlinear characteristics of the relationship between network tie quantity and innovation performance make it difficult for simplistic linear models to uncover the underlying association mechanisms.

(2)

Network Relationship Strength (NRS)

Most existing studies on the relationship between innovation network structure and innovation performance treat tie strength among nodes as a constant, overlooking its dynamic influence. However, varying levels of network tie strength significantly affect the efficiency of knowledge transfer between nodes, thus shaping collaborative innovation patterns [48]. Thomas Ritter demonstrated that strong network ties enable collaborating institutions to allocate resources more effectively and reduce barriers to knowledge transfer. However, research also indicates that weak ties play a crucial role during the initial phases of innovation. Such weak ties introduce heterogeneous perspectives that stimulate novel ideas. Conversely, strong ties with long-term partners may foster cognitive homogenization by excessively reinforcing specific innovation trajectories, thereby increasing the risk of evaluation biases such as “false positives” [49]. Furthermore, tighter interorganizational connections result in higher exit costs, which can weaken learning motivation and ultimately reduce innovation performance [50]. Wu et al. revealed that both strong and weak ties contribute distinctly to technological innovation performance via different resource mechanisms [51]. Additional studies suggest that the overall relationship between network tie strength and innovation performance exhibits complex nonlinear characteristics.

2.3. Configuration Method and Its Application Research

The fuzzy-set qualitative comparative analysis (fsQCA) method integrates the strengths of both quantitative research and qualitative analysis [52]. As shown in Table 1, traditional methods follow the physics paradigm [53] and marginal analysis techniques, adopting an atomistic perspective that emphasizes linear, symmetric causal relationships between individual antecedent conditions and outcomes. In contrast, fsQCA adopts a configurational perspective, emphasizing the analysis of configuration effects formed by interdependent organizational attributes, or antecedent conditions [9]. Unlike conventional methods such as multiple regression, factor analysis, and structural equation modeling—which infer causality from correlations among variables—fsQCA infers causality based on set-theoretic relationships [54]. Grounded in complex systems theory, fsQCA offers a nuanced analysis of complex management problems, addressing conjunctural causation, causal asymmetry, and multiple equifinal paths [11].

Conventional regression and correlation analyses mainly focus on investigating sufficient conditions while neglecting the logical inference of whether antecedent conditions are necessary for outcomes [55]. In reality, identifying necessary conditions is equally important for examining complex management issues, as they hold causal primacy over sufficient conditions [56], often referred to as “bottleneck conditions.”. To address this limitation, Dul proposed the necessary condition analysis (NCA) to identify necessary-but-not-sufficient conditions. NCA not only assesses whether an antecedent condition is a prerequisite for specific outcomes but also quantifies the effect size of necessary conditions by determining the minimum threshold required to achieve the target outcome [57]. This method thus serves as a robust complement to fsQCA.

Table 1. Comparison of the Mechanisms of Traditional Statistical Analysis Methods and fsQCA.

Analytical Perspective	Traditional Statistical Methods	fsQCA
Theoretical premise	Based on the reductionism hypothesis [53]	Grounded in complex systems theory [8]
Variable interactions	“Physical phenomena” between variables, following the physics paradigm [53]	“Chemical reactions” of interacting variables [9]
Causal complexity	“Net effect” of single factors and simple additivity	“Net effect” of single factors and simple additivity Complex phenomena involving multi-factor interactions, interdependence, and conjunctural causation [44]
Causal relationships	Symmetric, linear cause–effect relationships [58]	Asymmetric causality [59] and equifinal paths [11]
Analytical foundation	Correlational relationships between variables [54]	Set-theoretic relationships [9]

Table 1. Comparison of the Mechanisms of Traditional Statistical Analysis Methods and fsQCA.

Analytical Perspective	Traditional Statistical Methods	fsQCA
Theoretical premise	Based on the reductionism hypothesis [53]	Grounded in complex systems theory [8]
Variable interactions	“Physical phenomena” between variables, following the physics paradigm [53]	“Chemical reactions” of interacting variables [9]
Causal complexity	“Net effect” of single factors and simple additivity	“Net effect” of single factors and simple additivity Complex phenomena involving multi-factor interactions, interdependence, and conjunctural causation [44]
Causal relationships	Symmetric, linear cause–effect relationships [58]	Asymmetric causality [59] and equifinal paths [11]
Analytical foundation	Correlational relationships between variables [54]	Set-theoretic relationships [9]

In summary, although existing studies have examined factors influencing innovation collaboration from various perspectives, they lack integrated research on cross-level, multidimensional interactions among the innovation environment, innovation actors, and innovation networks, as well as their combined effects on collaborative performance. Moreover, prior research shows that most antecedent conditions exhibit significant nonlinear effects on outcomes, indicating that simple linear regression analysis alone cannot fully uncover their complex interactive mechanisms.

Rooted in systems thinking, fsQCA’s configurational analysis aligns with the complex systems perspective of innovation ecosystems and offers a more nuanced analysis of nonlinear interactions among multiple factors. Therefore, this study integrates the three dimensions of innovation environment, innovation actors, and innovation networks into a unified analytical framework grounded in innovation ecosystem theory. Following the configurational thinking paradigm and employing a combined fsQCA and NCA approach, the study overcomes previous single-level analyses of innovation collaboration by exploring the joint effects of cross-level, multi-factor interactions on collaboration performance, focusing on the interactive effects of multiple conditions and employing exploratory configurational identification.

3. Methods

3.1. Research Framework and Process

We adopt an integrated perspective of complex systems theory and innovation ecology to construct an “innovation environment–innovation actors–innovation networks” analytical framework grounded in the innovation ecosystem perspective. (as shown in Figure 1). Following the configurational paradigm, we employ fsQCA methodology to investigate nonlinear interactions among multiple factors influencing innovation collaboration performance.

Innovation ecosystem construction: Within the innovation ecosystem, cross-level linkages and multi-factorial coupling among innovation environments, actors, and networks generate multiple equifinal driving patterns that foster high innovation collaboration performance.

Indicator measurement: Based on the innovation ecosystem framework, key factors influencing innovation collaboration performance are identified and quantified using BERTopic topic modeling and Neo4j knowledge graph analysis.

Pattern identification: fsQCA and NCA methods are applied to identify multiple pathways leading to high innovation collaboration performance. The configurational results are systematically analyzed to explore their operational mechanisms and practical applications.

The following subsections detail the processes of innovation ecosystem construction and configurational analysis of collaboration performance in this study.

3.2. Theoretical Framework: Constructing the Innovation Ecosystem for Innovation Collaboration

We construct a collaborative innovation ecosystem model, as illustrated in Figure 2. The model centers on three core components—innovation environment, innovation actors, and innovation networks—with an emphasis on their synergistic interactions. The evolution of innovation ecosystems is driven by nonlinear coupling and co-evolutionary mechanisms between endogenous and exogenous forces, rather than by their linear superposition. As typical complex systems, collaborative innovation ecosystems exhibit self-organizing and adaptive characteristics through interactive evolution.

Entities such as research institutions, enterprises, and universities engage in collaborative interactions facilitated by these self-organizing and adaptive mechanisms. Each innovation collaboration organization constitutes a collaborative innovation ecological subsystem with specific functional structures. These subsystems maintain relative independence while forming multi-level, dynamically evolving complex innovation ecosystems via regional and cross-regional network connections.

From the perspective of innovation ecosystems, endogenous evolutionary forces refer to dynamics driven by the capabilities and behavioral patterns of innovation actors themselves. These forces reflect actors’ proactive agency in resource integration and collaborative interactions, which can internally propel transformations within the system [60]. In contrast, exogenous evolutionary forces arise from external constraints, such as macro-level institutions and geographic endowments. These forces guide the overall evolutionary direction and trajectory of the system by establishing rules and creating windows of opportunity [61]. Based on the classification of the “determinant source,” the innovation environment is categorized as an exogenous evolutionary force, while innovation actors are classified as an endogenous evolutionary force.

Innovation networks exhibit a dual nature. On the one hand, network tie strength reflects actors’ endogenous choices regarding partner selection and interaction intensity. On the other hand, network size captures characteristics of the external environment, such as regional resource endowments. Consequently, network attributes simultaneously embed both the endogenous behavior of actors and the exogenous characteristics of environmental resources. We therefore conceptualize the innovation network as a “structurally embedded driving force,” serving as a mediating mechanism that links endogenous and exogenous forces, facilitating key transmission and regulatory functions between them.

As illustrated in the proposed mechanism model, the coupling interactions among the three forces are clearly delineated. First, the innovation environment generates windows of opportunity through geographic proximity and policy-institutional arrangements, providing the external impetus for the initiation and diffusion of innovation activities. Second, the structurally embedded innovation network forms conduits for the transmission and integration of innovation elements, knowledge flows, and actor interactions. Specifically, tie strength determines the depth and stability of resource exchanges, while network size reflects the accessibility of resources provided by the external environment. Finally, within the constraints imposed by both the environment and network structure, innovation actors leverage their capabilities, knowledge structures, and collaborative orientations to proactively update and reconfigure innovation resources, thereby activating and reshaping collaborative networks.

The interaction of these three forces constitutes a bidirectional, feedback-driven evolutionary mechanism. Exogenous environmental factors shape opportunity boundaries, structural networks facilitate resource transmission, and endogenous actors drive network reconfiguration and knowledge creation. Their dynamic interplay promotes the co-evolution and complex adaptation of the innovation ecosystem, reflecting system-level characteristics of nonlinear evolution.

Within this innovation ecosystem, innovation collaboration is fundamentally a nonlinear interactive process among its components. Various elements influence collaboration performance not through isolated effects but through interdependent, cross-level coupling relationships that collectively shape outcomes. The innovation environment, innovation actors, and innovation networks achieve synergistic alignment via both complementary and substitutable interactions. This enables cross-level linkages and multi-factor coupling, producing multiple equifinal configurations. These distinct driving patterns arise not from individual factors but from complex interactions among multiple antecedent conditions. Such interactions form differentiated driving modes that collectively promote high innovation collaboration performance.

Guided by the aforementioned theoretical framework, we adopt an innovation ecosystem and complex systems perspective, following the configurational thinking paradigm to integrate three dimensions—innovation environment, innovation actors, and innovation networks—within a unified analytical framework. Employing NCA and fsQCA methodologies, we investigate how cross-level, multi-factor interactions influence innovation collaboration performance.

The configurational analysis reveals multiple equifinal pathways through the identification of multi-factor interactions, emphasizing that under different combinatorial patterns, various factors can collectively—though differentially—drive the achievement of high innovation collaboration performance. This approach fundamentally demonstrates that superior performance outcomes may emerge from distinct configurations of antecedent conditions, wherein alternative combinations of environmental, organizational and network factors can produce similarly effective results through varied causal recipes.

3.3. Configurational Analysis of Collaboration Performance

3.3.1. Variable Measurement

Based on the innovation ecosystem theoretical framework, we select seven variables from three dimensions: innovation environment, innovation actors, and innovation networks. Variable descriptions are presented in

Table 2. Variable Measurement and Description.

Variable Type		Variable Name (Abbreviation)	Description
Condition variables	Innovation environment	Geographic proximity (Geo)	Spatial distance between collaborating institutions
		Institutional proximity (Ins)	Degree of similarity in policy environments between collaborating institutions
	Innovation actors	Technological proximity (Tec)	Extent of technological similarity between collaborating institutions
		Collaboration tendency (Col)	Frequency of prior collaborative engagements
	Innovation networks	Network relationship quantity (NRQ)	Total number of direct collaborative ties within the innovation network
		Network relationship strength (NRS)	Intensity of existing collaborative ties within the innovation network
Outcome variable		Innovation collaboration performance (ICP)	The richness of innovative elements and technical details in collaborative patents

, followed by detailed analysis.

(1)

Outcome Variable

In this study, we use innovation collaboration performance as the outcome variable, with collaborative patents serving as tangible representations of institutional achievements. The quality of these patents effectively reflects collaborative performance. Patent knowledge breadth—defined as the complexity of knowledge embedded in patents—serves as a reliable proxy for patent quality [62]. Traditional approaches that rely solely on counting patent classification codes fail to capture intrinsic differences among codes [63]. Drawing on Zhang et al.’s methodology, we measure patent quality through knowledge breadth [62]. Specifically, IPC classification codes are extracted from the patent data, and weighted analyses are conducted at the subgroup level. The Herfindahl–Hirschman index (HHI) is applied to account for structural relationships among classification groups as shown in Equation (1).

$$ICP=1-\Sigma {\alpha }^2$$

(1)

Here, α denotes the proportion of each primary classification group within the total classification groups of collaborative patents. Higher performance values indicate that the patents incorporate more innovative elements and technical details, reflecting greater complexity and, in turn, higher innovation collaboration performance.

(2)

Condition Variables

• Geographic Proximity (Geo)

In this study, we calculate distances using the great-circle formula based on the latitude and longitude coordinates of collaborating institutions [64], To eliminate dimensional differences and improve interpretability, the normalization approach proposed by Yang et al. is adopted to convert distances into proximity values [65]. This transformation intuitively reflects the spatial closeness of collaborative relationships as shown in Equations (2) and (3).

$$D_{ij}=R\times arccos\left[cos\left(Y_i\right)\right]\times cos\left(Y_J\right)\times cos(X_i-X_j)+sin\left(Y_i\right)sin\left(Y_j\right)],$$

(2)

$${Geo}_{ij}=1-{\displaystyle \frac{D_{ij}}{\mathit{max}\left(D_{ij}\right)}},$$

(3)

Here, D_ij denotes the distance between institutions i and j; max (D_ij) is the maximum distance among all regional institution pairs; R denotes the Earth’s radius (R = 6371 km). X_i and X_j are the longitudes of institutions i and j, respectively, while Y_i and Y_j are their latitudes. The Geo_ij metric quantifies the geographical proximity between institutions i and j, with higher values indicating greater spatial closeness. For patents involving multiple collaborating institutions, the arithmetic mean of all pairwise proximity values is used as the overall proximity score.

• Institutional Proximity (Ins)

In this study, we measure institutional proximity based on the consistency of economic policies and the similarity of investment and business environments across different cities or regions [66]. Institutional proximity analyzes policy documents issued by multiple levels of government to reveal their external characteristics and internal logic [67]. This study leverages policy document data and employs BERTopic for thematic modeling. While latent Dirichlet allocation (LDA), an unsupervised model based on word-bag statistics, represents documents as probabilistic mixtures of latent topics and models topics as word distributions [68], it struggles to capture the nuanced semantics of words in different contexts. In contrast, BERTopic integrates BERT’s deep semantic comprehension, allowing it to identify richer semantic information in text. Compared to LDA, BERTopic captures more profound semantic nuances [69], making it more suitable for analyzing policy texts with their complex and formal linguistic structures. BERTopic generates context-aware document embeddings using BERT and reduces their dimensionality with the UMAP algorithm [70]. UMAP is a dimensionality reduction technique designed to preserve the original topological structure of the data when mapping it from high-dimensional to low-dimensional space. Next, clustering is performed using the HDBSCAN algorithm [71], followed by the extraction of representative feature words through the c-TF-IDF algorithm [72]. HDBSCAN, a density-based hierarchical clustering algorithm, automatically determines the number of clusters and identifies noise points (outliers). The c-TF-IDF method extends traditional TF-IDF from the document level to the class (cluster) level. Instead of identifying keywords for individual documents, it extracts representative terms for topics, which consist of clusters of documents. Finally, using the c-TF-IDF representation (i.e., the topic vector) for each topic, cosine similarity between policy topics is calculated, as shown in Equations (4) and (5).

$${\mathrm{c}-\mathrm{T}\mathrm{F}-\mathrm{I}\mathrm{D}\mathrm{F}}_{t,c}={\mathrm{T}\mathrm{F}}_{t,c}\times {\mathrm{I}\mathrm{D}\mathrm{F}}_t=f_{t,c}\times log\left(1+{\displaystyle \frac{m}{f_t}}\right),$$

(4)

$$Ins\left(p,q\right)={\displaystyle \frac{\sum _{k=1}^d p_k\times q_k}{\sqrt{\sum _{k=1}^d p_k^2}\cdot \sqrt{\sum _{k=1}^n q_k^2}}},$$

(5)

Here, f_t,c represents the frequency of term t in class c, m is the average number of words per class across all classes, f_t is the total frequency of term t across all classes, and p and q are two n-dimensional vectors (the vector representations of provinces in the topic space).

• Technological Proximity (Tec)

In this study, we assess technological proximity by first identifying patents independently filed by each collaborating institution, supplemented by co-filed patent data when independent applications are unavailable. Text data are extracted from the “title” and “technical effect phrases” fields, the latter being more concise than abstract fields and thus better suited for avoiding redundancy and non-technical content. Using Python’s Jieba library f(version 3.9.10) or text segmentation, we apply the Word2Vec model to generate semantically rich word vectors that effectively capture the similarity between technical terms in patent texts [73]. Technological distance between institutions is then measured using cosine similarity. For patents involving multiple collaborators, the final technological proximity score is derived by averaging all pairwise distance values, offering a comprehensive measure of technical alignment, as shown in Equation (6).

$$Tec(i,\mathrm{j})={\displaystyle \frac{\sum _{k=1}^n i_k\times j_k}{\sqrt{\sum _{k=1}^n i_k^2}\cdot \sqrt{\sum _{k=1}^n j_k^2}}},$$

(6)

Here, i and j represent the text feature vectors of the collaborating institutions, respectively. A Tec(i, j) value approaching 1 indicates a smaller technological distance between the institutions i and j, whereas a value approaching 0 indicates a larger technological distance.

• Collaboration Tendency (Col)

Prior collaborative interactions between institutions foster trust, which in turn increases the likelihood of future cooperation [40]. The frequency of previous collaborations provides a quantifiable proxy for this trust-based relationship. In this study, collaboration tendency is operationalized by counting the number of previous cooperative engagements between institutions.

• Network Relationship Quantity (NRQ)

Degree centrality quantifies the number of direct connections a node maintains within a network [74]. We construct and analyze the collaboration network using Neo4j, importing nodes and relationships into the Neo4j environment. The Cypher query language and Neo4j’s Graph Data Science (GDS) library modules are used to calculate the degree centrality of each collaborating institution. For a network containing g nodes, the degree centrality of node i is computed as shown in Equation (7):

$$NRQ={\sum }_{j-1}^g x_{ij},$$

(7)

Here, x_ij indicates whether a direct connection exists between node i and node j (1 if it exists, 0 otherwise), and i ≠ j. This formula calculates the total number of direct connections between node i and all other g-1 nodes in the network.

• Network Relationship Strength (NRS)

In link prediction tasks, the strength of potential ties is commonly measured by the number of common neighbors, with a higher count indicating stronger potential connections. However, common neighbors differ in their influence. The Adamic–Adar (AA) index assigns greater weights to common neighbors with lower degrees [75]. Common neighbors with fewer connections tend to be more “focused” on specific nodes or groups, thereby exerting more unique and precise influence on information flow and facilitating the formation of stronger, more targeted ties. High-quality research outputs arise when collaborating institutions operate within relatively sparse networks while maintaining strong ties [76]. Sparse network structures thus amplify the impact of strong ties [77]. Based on this, we employ the AA index to quantify network tie strength. The AA values between collaborating institutions are calculated using built-in modules from the Neo4j Graph Data Science (GDS) library as shown in Equation (8).

$$NRS(i,j)={\sum }_{z\in N\left(i\right)\cap N\left(j\right)} {\displaystyle \frac{1}{\mathrm{l}\mathrm{o}\mathrm{g}\left|N\right(z\left)\right|}}$$

(8)

Here, z denotes the set of common neighbors between nodes (institutions) i and j, and NRS (i, j) represents the Adamic–Adar coefficient between these nodes. A higher coefficient indicates a stronger tie strength between the two collaborating institutions.

3.3.2. Variable Calibration

Fuzzy-set Qualitative Comparative Analysis (fsQCA), based on set-theoretic principles, requires calibration of outcome and condition variables [54]. Following established approaches [6,78], we set the 95th percentile, the mean, and the 5th percentile of the data as the full membership threshold, crossover point, and full non-membership threshold, respectively. This calibration converts raw continuous variables into membership scores appropriate for fsQCA analysis,.as shown in

Table 3. Calibration Anchors for Set-Membership.

Variable Type	Variable Name		Code	Calibration Anchor Points
				Full Membership	Crossover Point	Full Non-Membership
				(μ = 0.95)	(Mean)	(μ = 0.05)
Outcome variable	Innovation collaboration performance		ICP	0.765	0.513	0
Condition variables	Innovation environment	Geographic proximity	Geo	0.999	0.963	0.850
		Institutional proximity	Ins	1.000	0.785	0.314
	Innovation actors	Technological proximity	Tec	0.864	0.527	0.279
		Collaboration tendency	Col	11.000	3.098	1.000
	Innovation networks	Network relationship quantity	NRQ	22.133	7.035	1.000
		Network relationship strength	NRS	15.870	3.663	0.910

.

4. Empirical Analysis

Natural Language Processing (NLP), a frontier field marked by rapid advancement and high innovation intensity, exhibits strong interdisciplinary characteristics. Patents in this domain often involve interactions across diverse technological domains and innovation elements, making collaborative patents a crucial indicator in quantitative evaluations of innovation outcomes [79]. Accordingly, we employ collaborative patents in the NLP field as proxies for measuring innovation performance. This approach facilitates the identification of complex factor interactions and configurational pathways, thereby uncovering the collaborative mechanisms driving innovation performance.

4.1. Sample Selection and Data Sources

The data sources for this study include patent data and regional AI policy documents from multiple Chinese provinces. The patent data were obtained from the IncoPat database (https://www.incopat.com) in February 2024. Based on the research topic, a set of search terms was developed to retrieve the relevant patents, which were then downloaded for subsequent analysis. Considering the rapid development of the NLP field and the time lag between patent application and publication, we selected Chinese collaborative patents filed between 2019 and 2023. Given the prevalence and representativeness of tripartite innovation systems (industry–university–research) and dual-entity systems (university–enterprise and research institute–enterprise collaboration [80]), we focus on these two innovation models. After excluding non-collaborative patents filed by a single institution and those with incomplete information, a total of 1473 patents were retained as the final case sample. To calculate technological proximity, we retrieved the patents independently filed by the collaborating institutions in the sample outside of their collaborative activities, obtaining a total of 56,019 patents.

Regional policy data were collected from the PKULaw platform (https://www.pkulaw.com). Given the limited availability of policy data, specifically on NLP, and recognizing that NLP is a key subfield of artificial intelligence, the study used “artificial intelligence” as the search term to retrieve relevant policies issued by Chinese provincial governments between 2019 and 2023, ensuring sufficient data coverage.

This study focuses on collaborative innovation ecosystems that span organizational and geographical boundaries. Unlike previous research, which often uses “regions” or “provinces” as case studies, we emphasize identifying and analyzing cross-regional innovation collaboration teams formed by diverse actors through synergistic cooperation within regional innovation ecosystems. Thus, this study not only extends research on regional innovation ecosystems but also reveals how innovation actors construct efficient and collaborative ecosystems through cross-regional linkages.

In specific technological innovation activities, a collaborative patent is co-created by multiple entities, such as universities, research institutes, and enterprises, forming a goal-aligned innovation collaboration team. Collaborative patents capture key fields, such as the types of participating entities and technological domains, providing a comprehensive view of the interaction patterns among innovation actors during the collaboration process. These interactions reflect underlying factors, such as resource allocation, knowledge structure, and relationship strength. Moreover, the standardized and structured nature of patent data enhances its reliability and validity in identifying innovation actor relationships, measuring collaboration behaviors, and constructing network structures. This provides a solid data foundation for research on collaborative innovation ecosystems. Therefore, the 1473 collaborative patents used in this study are not treated as a complete innovation ecosystem but as an analytical tool that captures the multidimensional information contained within them to depict the micro-ecosystems formed around collaborative innovation activities.

These data reflect real cross-organizational and cross-regional collaboration mechanisms in technological innovation and, compared to traditional studies focused on regional boundaries, align more closely with the cross-boundary collaboration logic this study aims to uncover.

4.2. Necessary Condition Analysis of Innovation Collaboration Performance

Before conducting configurational analysis, it is necessary to assess whether any conditions are essential for achieving high innovation collaboration performance. NCA offers two methodological approaches—Ceiling Regression (CR) and Ceiling Envelopment (CE)—to accommodate different data types and evaluate the necessity of antecedent conditions. CR is suitable for continuous variables and discrete variables with more than five levels, while CE applies to discrete variables with five or fewer levels and dichotomous variables.

Since all variables in this study are continuous, CR is used as the primary method, with CE applied for comparison. The NCA framework evaluates necessity by calculating effect sizes using Monte Carlo permutation tests. In addition, bottleneck analysis quantifies the minimum necessary levels of these conditions [81]. Compared to necessity testing in fsQCA, NCA not only identifies whether a condition is necessary but also quantifies the degree of necessity. A condition is considered necessary only if the effect size exceeds 0.1 and the Monte Carlo permutation test yields statistical significance (p < 0.05).

As shown in

Table 4. NCA of the Necessity of Single Conditions.

Variables	Method	Accuracy	Ceiling Zone	Scope	Effect Size (d)	p Value
Geographic proximity (Geo)	CR	100%	0.000	0.88	0.000	0.778
	CE	100%	0.000	0.88	0.001	0.754
Institutional proximity (Ins)	CR	100%	0.002	0.87	0.000	0.172
	CE	100%	0.000	0.87	0.001	0.179
Technological proximity (Tec)	CR	99.6%	0.002	0.91	0.002	0.411
	CE	100%	0.002	0.91	0.002	0.511
Collaboration tendency (Col)	CR	100%	0.000	0.88	0.000	1.000
	CE	100%	0.000	0.88	0.000	1.000
Network relationship quantity (NRQ)	CR	100%	0.003	0.89	0.003	0.044
	CE	100%	0.006	0.89	0.006	0.029
Network relationship strength (NRS)	CR	100%	0.000	0.88	0.000	1.000
	CE	100%	0.000	0.88	0.000	1.000

Note: Effect size (d): 0.0 ≤ d < 0.1 indicates “low level”; 0.1 ≤ d < 0.3 indicates “medium level”; p value: permutation test in NCA with number of resamples = 10,000.

, none of the examined variables—geographic proximity, institutional proximity, technological proximity, collaboration tendency, or network relationship strength—met the criteria (all p > 0.05). Although network relationship quantity reached statistical significance (p < 0.05), its effect size remained below the 0.1 threshold. Therefore, none of these factors constitute necessary conditions for achieving high innovation collaboration performance.

Table 5. NCA of the Necessity Bottleneck Level (%) of Single Conditions.

Innovation Collaboration Performance (ICP)	Geographic Proximity (Geo)	Institutional Proximity (Ins)	Technological Proximity (Tec)	Collaboration Tendency (Col)	Network Relationship Quantity (NRQ)	Network Relationship Strength (NRS)
0	NN	NN	NN	NN	NN	NN
10	NN	NN	NN	NN	NN	NN
20	NN	NN	NN	NN	NN	NN
30	NN	NN	NN	NN	NN	NN
40	NN	NN	NN	NN	0.0	NN
50	NN	NN	NN	NN	0.2	NN
60	NN	NN	NN	NN	0.4	NN
70	NN	NN	NN	NN	0.5	NN
80	NN	NN	NN	NN	0.7	NN
90	NN	NN	NN	NN	0.9	NN
100	5.3	7.1	8.9	NN	1.0	NN

Note: CR method; NN = not necessary.

further illustrates the minimum condition levels required to achieve specific outcome thresholds [52]. As shown in Table 5, achieving 80% innovation collaboration performance requires attaining 70% of the network relationship quantity level. No bottleneck levels were identified for other conditions, indicating that they do not impose minimum thresholds for achieving specific performance outcomes.

Table 6. Results of the fsQCA Analysis of the Necessity of Single Conditions.

Antecedent Variables	High Innovation Collaboration Performance		Low Innovation Collaboration Performance
	Consistency	Coverage	Consistency	Coverage
Geographic proximity (Geo)	0.769	0.663	0.806	0.475
~Geographic proximity (~Geo)	0.392	0.747	0.430	0.560
Institutional proximity (Ins)	0.667	0.646	0.715	0.560
~Institutional proximity (~Ins)	0.458	0.701	0.467	0.490
Technological proximity (Tec)	0.574	0.742	0.665	0.589
~Technological proximity (~Tec)	0.682	0.748	0.709	0.533
Collaboration tendency (Col)	0.344	0.734	0.444	0.649
~Collaboration tendency (~Col)	0.835	0.687	0.818	0.460
Network relationship quantity (NRQ)	0.428	0.715	0.523	0.597
~Network relationship quantity (~NRQ)	0.759	0.699	0.751	0.473
Network relationship strength (NRS)	0.364	0.754	0.477	0.676
~Network relationship strength (~NRS)	0.844	0.702	0.827	0.471

Note: ‘~’represents logical “NOT”.

reports the necessity analysis results for individual conditions based on the fsQCA conducted in this study. All consistency scores for high innovation collaboration performance fall below the 0.9 threshold [50] suggesting that no single antecedent condition has a sufficiently strong and independent effect to qualify as necessary for the outcome. These findings are consistent with the results of the NCA. Additionally, analysis of the X-Y scatter plot between the condition variables and the outcome variable revealed that approximately one-third of the case points fall above the diagonal. This pattern suggests that the condition variable cannot be regarded as a necessary condition for achieving high performance [82]. A detailed illustration is provided in the Appendix (Figure A2, Figure A3, Figure A4, Figure A5, Figure A6 and Figure A7).

The above findings further confirm that innovation collaboration is inherently a nonlinear, interactive process involving multiple actors and factors within an innovation ecosystem, where performance emerges from coupled interactions across multiple levels. No single condition alone can determine high innovation collaboration performance. Reductionist analytical approaches that assess elements in isolation—typically through “net effect” analyses—are limited in explaining cross-level innovation synergies and fail to capture multiple conjunctural causes.

4.3. Configurational Analysis Results of Innovation Collaboration Performance

4.3.1. Analysis of High-Performance Configurations

We used fsQCA 3.0 software to conduct a configurational analysis of high innovation collaboration performance. The frequency threshold was set to 1 to retain 75% of the original sample, and the consistency threshold was set at 0.85. The PRI (Proportional Reduction in Inconsistency) threshold was set at 0.7 to avoid the “divergent outcomes from the same cause” phenomenon, where a configuration could simultaneously lead to both the outcome and its negation. After fsQCA computation, the intermediate solution was primarily used, with the parsimonious solution as a supplement, to identify core and peripheral conditions for pathway analysis, the relevant truth table is shown in the appendix (Figure A1). The detailed results are shown in

Table 7. Configuration Analysis of High Innovation Collaboration Performance.

Antecedent Condition	H1	H2	H3	H4	H5
Geographic proximity (Geo)	⬤	$\otimes$	⬤	●
Institutional proximity (Ins)	⊗	⬤		●	⊗
Technological proximity (Tec)			⬤	⬤	⊗
Collaboration tendency (Col)	⊗		⬤	⊗	⬤
Network relationship quantity (NRQ)		●	●	⊗	●
Network relationship strength (NRS)	$\otimes$	⊗	⊗	⬤
Consistency	0.891	0.955	0.939	0.931	0.940
Raw coverage	0.230	0.127	0.154	0.174	0.131
Unique coverage	0.029	0.005	0.013	0.052	0.007
Overall solution consistency	0.833
Overall solution coverage	0.460

Note: “⬤” represents the presence of a core condition, “⊗” represents the absence of a core condition, “●” represents the absence of a peripheral condition, “$\otimes$” represents the absence of a peripheral condition, and a blank indicates that the presence or absence of the condition has no effect on the outcome variable.

.

The configurations in Table 7, despite minor differences in raw and unique coverage, consistently contribute to high innovation collaboration performance. Together, these configurations constitute sufficient conditions for high performance. Unlike traditional linear analytical approaches that seek a single optimal solution, fsQCA explores the complex mechanisms of interacting factors to identify multiple equifinal pathways to the same outcome. This illustrates the principle of “equifinality,” where different configurations function as equally effective pathways to high performance.

The configurational analysis in this study yields an overall solution coverage score of 0.833, indicating strong explanatory power of the model. With an overall solution consistency of 0.460, the results explain approximately 46% of the factors contributing to high innovation collaboration performance. We follow the configurational theorizing approach of fsQCA [76]. By systematically analyzing the configurations composed of five core conditions influencing innovation cooperation performance, we explore the synergistic effects and underlying logical relationships among these variables. Based on substantive theoretical reasoning, we identify four archetypal patterns that effectively drive high innovation cooperation performance: Environment-Dominant Pattern, the Innovation Environment–Innovation Actor Synergistic Pattern, the Innovation Actor–Innovation Network Dual-Driven Pattern, and the Innovation Actor-Dominant Pattern.

(1)

Innovation Environment-Dominant Pattern

This configuration is defined by two core conditions: geographical proximity and institutional proximity. Path H1 is primarily driven by geographical proximity, whereas Path H2 is driven by institutional proximity. Both paths therefore fall within the Innovation Environment–Dominant Type, although they differ in their contextual conditions and operational mechanisms. Geographical and institutional proximity jointly create a favorable environment for collaborative innovation by lowering communication and coordination costs and enabling actors to advance joint activities within a shared institutional framework.

In configuration H1, geographic proximity is the core condition, whereas institutional proximity and collaboration tendency are absent; the absence of network relationship strength functions as a peripheral condition. Geographic proximity provides an effective medium for cross-institutional knowledge exchange by minimizing spatial distance. This spatial advantage facilitates faster diffusion of tacit knowledge and systematic integration of explicit knowledge, compensating for deficiencies in collaborative experience. Furthermore, low institutional proximity—indicating significant heterogeneity in policy focus areas among collaborators—helps mitigate inter-organizational competition.

In this configuration, the collaboration between the Shanghai Jiao Tong University School of Medicine and Suzhou Fubian Medical Technology Co., Ltd. serves as a representative example. The two parties had only one previous collaboration, representing a typical “weak cooperation foundation” scenario. However, the differences in policy environments between the two provinces created complementary advantages for the collaboration. Shanghai’s artificial intelligence policies focus more on healthcare and regulatory standards, while Jiangsu Province has vigorously promoted AI development in education and technological applications. This policy divergence reduced pressure from homogeneous competition, while providing multidimensional policy support in areas such as technology R&D, talent development, and application scenarios. These external factors strengthened cooperation incentives. Additionally, the geographical proximity of the two regions allowed joint R&D with lower communication costs, thereby improving collaboration efficiency and output performance. Additional cases can be found in Appendix A.1 H1 Case Interpretation.

In configuration H2, the absence of institutional proximity and network relationship strength functions as core conditions, while geographic proximity and the absence of network relationship quantity serve as complementary conditions. When collaborating entities operate within a governance framework characterized by convergent policy priorities, aligned strategic orientations and development paths can effectively reduce conflicts in developmental focus, thereby improving collaborative efficiency. Institutional alignment provides a stable foundation for cross-regional collaboration, where policy convergence supports consensus-building in industrial support and resource allocation. Supported by network relationship quantity, institutional proximity enhances the social influence of collaborating institutions and enables broader access to shared resources. Although the absence of strong network ties may hinder knowledge transfer and interaction, institutional convergence fosters standardized operational frameworks that mitigate the adverse effects of weak ties on collaboration quality.

Typical cases of this configuration include the collaboration between Changsha University of Science and Technology and Zhengzhou Xinda Advanced Technology Research Institute, and the cooperation between Qingdao Intelligent Industry Technology Research Institute and State Grid Zhejiang Electric Power Co., Ltd. The network relationship strength in both collaborations is only 1.442, significantly lower than the sample average of 3.663, indicating weak prior interactions and limited network embedding between the innovation actors. However, despite the weak network foundation, these two groups successfully established cooperative relationships. The key reason lies in the institutional support provided by the high consistency of the cross-regional policy environment.

Specifically, Hunan Province, where Changsha University of Science and Technology is located, and Henan Province, where Zhengzhou Xinda Advanced Technology Research Institute is based, share common features in their artificial intelligence policy frameworks. These include the development of intelligent application scenarios for enterprises, the deepening of the AI education system, and the innovative application of smart healthcare. The multidimensional coverage of these policies not only reduced institutional friction in cross-regional collaboration but also provided joint institutional incentives for cooperation through mechanisms such as research project funding and the co-building of innovation platforms. Additional cases are detailed in Appendix A.2 H2 Case Interpretation.

(2)

Innovation Environment–Innovation Actor Synergistic Pattern

In configuration H3, geographic proximity, technological proximity, collaboration tendency, and the absence of network relationship strength function as core conditions, while network relationship quantity serves as a complementary condition. This configuration indicates that when collaborating entities benefit from both geographical and technological advantages, along with prior collaboration experience, they can still improve innovation performance through multidimensional complementary mechanisms, even in the presence of weak network relationship strength. Geographic proximity reduces spatiotemporal barriers to tacit knowledge transfer, offering a physical channel for frequent knowledge exchange and technical collaboration. Technological proximity ensures that both parties share similar knowledge structures and research domains. Furthermore, accumulated collaborative experience fosters organizational tacit knowledge and reinforces codified explicit knowledge.

A typical case of this configuration involves collaborative innovation among the Institute of Automation, Chinese Academy of Sciences, State Grid Tianjin Electric Power Company, China Electric Power Research Institute, and State Grid Corporation of China. These four institutions are located in Beijing and Tianjin, geographically close, providing convenient proximity for cross-institutional collaboration and reducing communication and coordination costs. Although the institutions have collectively participated in nine joint projects or technical collaborations, their network relationship strength is only 1.820, significantly lower than the overall sample average of 3.663. This suggests that despite frequent collaboration, the depth of their connections remains weak, characteristic of a “weak network—multiple historical interactions” scenario. Notably, these institutions share technological proximity in areas such as system-on-chip devices, electronic equipment, and storage media, following similar technological paradigms and engineering implementation paths. This shared knowledge structure and technical logic enable smooth collaboration, even without deep network embedding. Additional cases are detailed in Appendix A.3 H3 Case Interpretation.

(3)

Innovation Actor–Innovation Network Dual-Driven Pattern

Configuration H4 includes technological proximity, absence of collaboration tendency, absence of network relationship quantity, and network relationship strength as core conditions, with geographic and institutional proximity as complementary conditions. This pattern indicates that when collaborating actors are positioned peripherally in the network and lack prior cooperative experience, the combination of technological proximity and strong-tie network relationships can overcome structural disadvantages and drive significant innovation performance improvements. In innovation networks, edge-positioned institutions face restricted access to information and resources because of limited direct connections. However, strong alignment in technological trajectories reduces collaboration barriers and resource dependency. Furthermore, strong-tie networks build trust foundations, enabling stable collaboration channels that offset higher coordination costs arising from lack of prior experience.

A typical case of this configuration is the collaboration between Zhengzhou Qingda Industrial Technology Research Institute and Zhengzhou University of Light Industry. The two parties have engaged in only three joint R&D activities, indicating relatively few collaborations, with a network relationship strength of 6.5, below the sample average of 7.035. This suggests that the external resources accessible to the collaboration are limited, and the network coverage is low. Notably, the two institutions have a high degree of alignment in their technical research directions, both focusing on cutting-edge intelligent technologies such as knowledge graph construction, deep learning algorithms, and multimodal data processing, creating significant technological proximity. Their collaboration network relationship strength is 4.3, higher than the overall average, indicating that, despite limited network resources, the connection between the two parties is closer and the quality of interaction is higher. This “technological convergence—strong connection” structure compensates for the disadvantage of limited resource access, providing strong internal support for collaborative innovation. Additional cases are detailed in Appendix A.4 H4 Case Interpretation.

(4)

Innovation Actor-Dominant Pattern

In configuration H5, the absence of institutional and technological proximity, along with collaboration tendency, serve as core conditions, while network relationship quantity functions as a complementary condition. This configuration shows that despite dual heterogeneity constraints in policy systems and technological domains, high innovation collaboration performance can be achieved through accumulated cooperative experience and advantages in network connection quantity. Long-term collaboration fosters trust capital and experiential norms, forming informal coordination mechanisms that overcome institutional differences. Meanwhile, increased network relationship quantity compensates for institutional and technological disparities by integrating information resources and enhancing actors’ adaptability to heterogeneous environments.

A typical case of this configuration is the collaboration between Tencent Technology (Shenzhen) and Beijing Jiaotong University. The two institutions are located in Guangdong Province and Beijing, with differing artificial intelligence policy focuses: Beijing’s policies emphasize enterprise technological innovation and AI education, while Guangdong’s policies support enterprises, education, and areas like smart healthcare and legal technology.

As a result, the policy support fields do not entirely overlap, indicating institutional divergence. Additionally, Tencent Technology focuses on intelligent question-answering methods, image semantic analysis, and voice recognition, while Beijing Jiaotong University specializes in building drug–disease relationship networks and processing judicial documents. The technological proximity is low, with clear differences in their knowledge backgrounds, algorithmic systems, and application scenarios. Despite the incomplete alignment of policy orientations and low technological proximity, the two parties have built high mutual trust and effective collaboration through 11 prior collaborations. This has enabled them to overcome differences in external policies and technological foundations, creating an intrinsic driving force for sustained cooperation. Additional cases are detailed in Appendix A.5 H5 Case Interpretation.

Comparative analysis of innovation collaboration patterns reveals that innovation actors act as core drivers in three of the four identified configurations, highlighting their pivotal role as endogenous evolutionary forces within collaborative innovation ecosystems. The agency of innovation actors—rooted in inherent human creativity and the fundamental pursuit of value maximization—emerges as a critical catalyst driving innovation collaboration. This endogenous force not only facilitates transformative advances in actors’ innovative capabilities but also enhances the overall performance of the innovation ecosystem. Furthermore, under specific capability endowments of innovation actors, a substitutive relationship arises between the innovation environment and innovation networks. When collaborating entities exhibit technological proximity or substantial cooperative experience, they can achieve high innovation performance regardless of whether their comparative advantage lies in environmental factors (e.g., geographic or institutional proximity) or network characteristics (e.g., relationship strength or quantity). This functional equivalence indicates that actor capabilities fundamentally mediate ecosystem performance pathways.

4.3.2. Analysis of Low Performance Configurations

In the analysis of low-performance pathways, we found that the PRI consistency of the low-performance solution set was below 0.5. This indicates that the path lacks strong support in distinguishing between “high performance” and “low performance”. According to the standards of the QCA method, the PRI consistency threshold for configuration analysis should not fall below 0.5 to ensure the stability and validity of the configurations. Therefore, a PRI consistency below 0.5 suggests that the configuration path does not possess sufficient discriminative power and cannot be further analyzed as a valid low-performance configuration. This finding highlights the asymmetric nature of causality. In symmetric causal logic, if a particular configuration consistently leads to high performance, its “symmetric counterpart”—the logical inverse—should also consistently lead to low performance. In such cases, the PRI consistency of the low-performance solution set should theoretically remain high (typically above 0.5), allowing for a clear distinction between high and low performance cases. However, our analysis shows that the PRI consistency of the low-performance pathways is significantly low, suggesting that configurations leading to low performance do not merely represent the logical inverse of high-performance configurations. As a result, there is no symmetric causal mapping between high and low performance. This phenomenon further emphasizes the complexity and asymmetry of causal relationships in innovation research and reinforces the necessity of adopting a configurational perspective and utilizing the QCA method when analyzing innovation performance pathways.

4.4. Configurational Patterns for Achieving High Innovation Collaboration Performance

By conducting configurational analysis on factors influencing innovation collaboration performance, we identify distinct patterns that enhance performance and lead to high innovation collaboration outcomes. To systematically elucidate the operational effects of these driving patterns and analyze their application strategies, we develop a conceptual framework (Figure 3) illustrating the pathways to achieve high innovation collaboration performance. This framework consists of the following: (1) foundational conditions for high performance, (2) driving patterns formed by multi-factor interactions, and (3) the operational effects generated by each pattern.

As shown in Figure 3, different driving pattern types require specific combinations of antecedent conditions, forming differentiated configurational pathways that ultimately facilitate high innovation collaboration performance. Here, solid lines represent core conditions, while dashed lines represent peripheral conditions.

Pattern 1: Innovation Environment-Dominant Pattern, referring to the driving strategy of “Regional Agglomeration & Policy-Driven Advancement”.

Geographic proximity strengthens innovation agglomeration by reducing spatial barriers, thereby accelerating knowledge and resource flows. Concurrently, institutional proximity—manifested through aligned regional policy priorities—enables collaborators to reach strategic consensus and allocate resources synergistically. In this context, the innovation environment shifts from a passive backdrop to an active driver of collaborative innovation and value co-creation, systematically enhancing innovation performance.

Pattern 2: Innovation Environment–Innovation Actor Synergistic Pattern, which involves the strategy of “Regional Technological Competitiveness-Driven Advancement” This pattern highlights the synergistic integration of geographic and technological proximities. By precisely identifying technologically aligned actors within geographic clusters, it strengthens regional technological competitiveness and fosters the formation of a “regional technology ecosystem.” This framework facilitates coordinated innovation among regional actors, thereby driving enhanced innovation collaboration performance.

Pattern 3: Innovation Actor–Innovation Network Dual-Driven Pattern, characterized by the collaborative strategy of “Market Exclusivity of Technological Outcomes-Driven”.

This strategy utilizes stable collaborative relationships to mitigate competitiveness erosion caused by technology spillovers, thereby enhancing the market monopoly potential of technological innovations. Furthermore, the embeddedness of strong network ties reduces free-rider behavior, enhances collaboration stability and sustainability, and fosters deeper development of innovation partnerships.

Pattern 4: Innovation Actor-Dominant Pattern, focusing on the collaborative strategy of “Synergistic Advancement through Internalized Collaboration Experience & External Resource Absorption”.

This strategy positions collaborative history as the core of stability, with network resources acting as complementary factors that provide expansion momentum. By internalizing past collaborative experience and integrating external resource absorption, it promotes broader synergistic innovation and expands the scope of collaborative efforts.

4.5. Decision Tree-Based Selection of Innovation Collaboration Configurations and Contextual Matching

Although configurational analysis highlights differences among various configurations, there is still a lack of systematic guidance on how to select the most suitable configuration based on organizational characteristics and environmental conditions. To address this gap, this study proposes a context-dependent decision-tree framework to guide innovation actors in scientifically selecting the most appropriate collaboration configuration, based on factors such as their capabilities and the surrounding environment. This framework simplifies the complex decision-making process of configuration selection into clear and actionable paths, offering practical recommendations for organizations of various types.

Figure 4 presents the “context-dependent decision tree framework” developed in this study, based on five configurations (H1–H5). This framework begins by evaluating innovation actors’ capabilities and progressively assesses technological foundations, collaboration experience, environmental conditions, and network structure, providing actionable collaboration strategy paths for different types of organizations.

The decision tree begins in Figure 4 with the “innovation actor capability assessment,” focusing on two key factors: technological proximity and prior collaboration experience. When both organizations exhibit high consistency in technological direction and extensive collaboration experience, the knowledge absorption and communication costs between them are relatively low, leading to a high-matching configuration, H3. If either technological proximity or collaboration experience is lacking, or both are absent, further analysis of the environmental layer is required.

In the second layer, the decision tree moves to “external environment analysis,” assessing whether geographic proximity and policy support provide favorable conditions for collaboration. If the partners are located in nearby regions or share similar policy orientations, the innovation environment can partially compensate for gaps in technology and experience, enabling effective collaboration based on external conditions.

Based on the environmental assessment, the decision tree divides external factors into two categories: “environmental support sources” and “network relationship structure.” If the collaborative advantage primarily stems from geographic proximity, the path leads to H1; if driven by policy support, the path corresponds to H2. When technological foundations and experience need supplementation by network structure, different strategies are chosen based on the characteristics of the network relationships. If network tie strength and prior collaboration experience are particularly advantageous, organizations may choose H4 or H5 as their collaboration paths.

Overall, this decision tree framework clarifies the formation mechanisms of collaboration paths in different contexts, effectively converting abstract theoretical configurations into a strategic decision-making process that can be practically implemented by innovation actors.

4.6. Analysis of Innovation Collaboration Contexts and Industry Fit

To address the potential limitations of a homogeneous sample, this study constructs a “context type-configuration path” matrix (

Table 8. “Context Type-Configuration Path” matrix.

Context Type	Characteristic Description	Matching Configurations	Applicable Industries
(A) Innovation environment-dominant pattern	Innovation activities rely heavily on geographical agglomeration, policy convergence, and unified governance frameworks; highly sensitive to changes in the innovation environment.	H1/H2	Biopharma, artificial intelligence: strong regulatory regimes; approval and policy frameworks largely determine technological directions. Clean energy: strongly driven by the national “dual-carbon” strategy.
(B) Innovation environment–innovation actor synergistic pattern	Requires a combination of technological proximity, geographical proximity, and accumulated collaboration experience.	H3	Clean energy, artificial intelligence: rapid evolution of technological trajectories. semiconductors: highly dependent on collaborative R&D;
(C) Innovation actor–innovation network dual-driven pattern	High technological alignment + strong-tie networks; substantial geographical or policy heterogeneity across actors.	H4	Equipment manufacturing: typical strong-tie supply-chain structures. Chemical materials: stable upstream–downstream application networks. Semiconductors: supply chains are globally dispersed but technological paths remain highly locked-in.
(D) Innovation actor-dominant pattern	Significant technological and institutional heterogeneity, but strong collaboration experience and rich relationship networks compensate for coordination gaps.	H5	Equipment manufacturing: long-term collaboration experience shapes innovation processes. Chemical materials: relies on long-term process expertise and iterative experimentation networks.

) to enhance the external validity and cross-industry applicability of the findings. This matrix maps the four configuration types to various industry contexts based on factors such as technological attributes, supply chain structure, policy sensitivity, and collaboration modes, reflecting the cross-context extendability of the fsQCA results.

The adaptability of configuration types varies significantly across technological industries, especially in sectors that are policy-driven or undergo rapid technological iteration. The clean energy industry is heavily influenced by macro policies, such as the “dual-carbon strategy,” and is undergoing rapid technological evolution (e.g., hydrogen energy, energy storage). This industry relies on policy convergence, a stable institutional environment, and technical collaboration between actors, as well as experience complementarity. As a result, it aligns with both the innovation environment-dominant (context A) and innovation environment–innovation actor synergistic pattern (context B) configurations. Similarly, the emerging artificial intelligence industry is highly sensitive to data governance, algorithm regulation, and industry-leading policies, while also relying on cross-actor collaboration in algorithms, computing power, and application scenarios. This dual dependence on “policy-driven and technological collaboration” makes it suitable for both context A and context B frameworks.

In contrast, the semiconductor industry, despite its globally dispersed supply chain, exhibits a significant technological lock-in, with strong dependencies between R&D and manufacturing processes. It requires cross-module synchronous iteration and extensive tacit knowledge exchange, making it highly dependent on geographic proximity, technological proximity, and collaboration experience. Thus, it aligns with the innovation environment–innovation actor synergistic pattern (context B) configuration. Moreover, its highly integrated supply chain network exhibits a clear innovation actor–innovation network dual-driven pattern (context C).

Compared to industries with rapid technological updates, traditional sectors, such as equipment manufacturing and chemical materials, have more stable technological paths and long-established supply chain relationships. Their innovation primarily relies on strong relationship networks, such as “main manufacturers–suppliers demand side — supply side” or “raw materials-application ends,” and long-term accumulated cooperation experience, making them more aligned with the innovation actor–innovation network dual-driven pattern (context C) and innovation actor-dominant pattern (context D) configurations.

Overall, differences in policy sensitivity, technological coupling, supply chain structure, and collaboration experience across industries shape the dominant logic of their innovation collaboration. This leads to differentiated applicability of configuration paths across industry contexts.

In conclusion, by constructing the “context type-configuration path-industry fit” matrix, this study extends the original fsQCA results into a cross-industry contextual analysis framework, enhancing the external validity of the research. It also provides further explanatory multi-context theoretical support for innovation collaboration mechanisms in different technological fields.

4.7. Robustness Analysis

(1) We conduct two robustness tests by adjusting the consistency threshold and the case frequency cutoff. As shown in

Table 9. Robustness Test Results.

Antecedent Condition	Adjusting the Consistency Threshold					Modifying the Case Frequency Threshold
	H1	H2	H3	H4	H5	H1	H2	H3	H4	H5
Geographic proximity (Geo)	⬤	$\otimes$	⬤	●		⬤	⊗	⬤	●
Institutional proximity (Ins)	⊗	⬤		●	⊗	⊗	⬤		●	⊗
Technological proximity (Tec)			⬤	⬤	⊗			⬤	⬤	⊗
Collaboration tendency (Col)	⊗		⬤	⊗	●	⊗		⬤	⊗	⬤
Network relationship quantity (NRQ)		●	●	⊗	●		●	●	⊗	●
Network relationship strength (NRS)	$\otimes$	⊗	⊗	⬤		$\otimes$	$\otimes$	⊗	⬤
Consistency	0.891	0.955	0.939	0.931	0.940	0.891	0.955	0.939	0.931	0.940
Raw coverage	0.230	0.127	0.154	0.174	0.131	0.230	0.127	0.154	0.174	0.132
Unique coverage	0.029	0.005	0.013	0.052	0.007	0.014	0.005	0.013	0.052	0.011
Overall solution consistency	0.833					0.833
Overall solution coverage	0.460					0.460

Note: “⬤” represents the presence of a core condition, “⊗” represents the absence of a core condition, “●” represents the absence of a peripheral condition, “$\otimes$” represents the absence of a peripheral condition, and a blank indicates that the presence or absence of the condition has no effect on the outcome variable.

, the configurations remain entirely unchanged when the consistency threshold is increased from 0.5 to 0.8. When the case frequency threshold is adjusted from 1 to 2, the resulting solution pathways exhibit fundamental consistency with the original configurations, with only slight variations in consistency and coverage scores. These variations are insufficient to alter the substantive interpretation of the original configurations [54], confirming the robustness of the findings.

(2) For instance, the global COVID-19 pandemic that began in 2020 reshaped not only the global economy and social structure but also had a profound impact on the collaborative relationships and innovation behaviors of various actors. To systematically assess the specific effects of such external shocks on innovation collaboration models, this study divides the research sample into two key phases for comparative analysis: Phase 1 (2019–2021), which represents the pre-pandemic and early-pandemic period, and Phase 2 (2022–2023), which represents the post-pandemic and late-pandemic period, when society began to adapt and seek transformation.

By segmenting the study into these phases, we aim to identify changes in innovation collaboration behaviors started during the pandemic. Through configurational analysis, it compares the consistency of collaboration models across these periods, seeking to uncover the evolutionary pathways and internal mechanisms of innovation ecosystems in response to external shocks. This perspective not only highlights the complex dynamics of innovation collaboration but also underscores the limitations of traditional regression methods in capturing multi-factor interactions, thereby demonstrating the value of configurational analysis in understanding the dynamic evolution of innovation systems.

As shown in

Table 10. Comparative Analysis of Configurations for Collaborative Innovation Performance.

Antecedent Condition	L1a	L1b	L2	P1	P2	P3
Geographic proximity (Geo)	●	$\otimes$	⬤	⬤	⬤	●
Institutional proximity (Ins)	⊗	⊗	⊗	⊗	●	⬤
Technological proximity (Tec)	$\otimes$	●	⬤			●
Collaboration tendency (Col)	⬤	⬤	⊗	●	⊗	⬤
Network relationship quantity (NRQ)	⊗	⊗	●	●	$\otimes$
Network relationship strength (NRS)	$\otimes$	●	$\otimes$		●	⊗
Consistency	0.971	0.968	0.951	0.948	0.921	0.933
Raw coverage	0.118	0.127	0.140	0.121	0.204	0.170
Unique coverage	0.009	0.021	0.036	0.007	0.003	0.004
Overall solution consistency	0.939			0.793
Overall solution coverage	0.177			0.789

Note: “⬤” represents the presence of a core condition, “⊗” represents the absence of a core condition, “●” represents the absence of a peripheral condition, “$\otimes$” represents the absence of a peripheral condition, and a blank indicates that the presence or absence of the condition has no effect on the outcome variable.

, in the pre-pandemic innovation collaboration model, the cooperation logic of configurations L1a and L1b was similar to the original configuration H5, both emphasizing the cumulative effect of past collaboration history. The L2 configuration, on the other hand, achieved high-performance collaboration through the dual support of geographic proximity and technological similarity, suggesting that spatial proximity and aligned technological paths created favorable conditions for innovation collaboration.

In the post-pandemic phase, the innovation collaboration model underwent significant transformations. The P1 configuration retained the innovation environment-dominant logic of the original H1 configuration, continuing to emphasize the critical role of the external innovation environment in supporting collaboration. However, the P2 configuration, while still dependent on the environment, placed increased emphasis on the strength of network relationships. In times of uncertainty, close partner relationships became a key driver of high-performance outcomes. The P3 configuration highlighted the combined impact of institutional support and collaboration experience, demonstrating that, under changing external conditions, the integration of institutional safeguards and established collaboration models provides adaptability and resilience.

Overall, while the post-pandemic collaboration model retains elements of the original logic, it places greater emphasis on institutional safeguards, the stability of relational networks, and the dynamic adaptability of the innovation environment.

5. Conclusions and Prospects

5.1. Main Conclusions

We developed a collaborative innovation ecosystem encompassing the innovation environment, innovation actors, and innovation networks. Grounded in innovation ecosystem theory, we integrated multiple factors into a unified framework and employed configurational analysis using patent collaboration data from the Natural Language Processing (NLP) field. Our analysis systematically revealed the inter-level, multidimensional linkages within the ecosystem.

In this system, the enhancement of collaborative innovation performance was driven not by any single factor—such as the innovation environment, innovation actors, or innovation networks—but by the synergistic effects of multiple elements. Unlike traditional analyses that sought a single optimal path to high innovation collaboration performance, our configurational analysis identified multiple equifinal paths to high innovation performance. These paths were categorized into four innovation-driven models: Innovation Environment-Dominant, Innovation Environment–Innovation Actor Synergistic, Innovation Actor–Innovation Network Dual-Driven, and Innovation Actor-Dominant.

In three of these models, innovation actors served as the core driving force, acting as endogenous evolutionary drivers. The comparative analysis revealed that, under specific conditions of innovation actor capabilities, an equivalent substitution relationship existed between the innovation environment and innovation networks. Additionally, the decision tree framework we developed helped innovation actors, based on their capabilities and contextual factors, select the most suitable collaboration configuration.

To address the limitations of NLP-specific data, we developed a “context-type–configuration path” matrix to map the four configurations. Our study found that differences in policy sensitivity and supply chain structures across industries led to varying applicability of configuration paths in different industrial contexts. At the theoretical level, we provide a comprehensive comparison between fsQCA and traditional analytical methods across multiple dimensions—including theoretical premises, variable interactions, and causal complexity—thus establishing a solid methodological foundation. Grounded in innovation ecosystem theory, we developed a multidimensional analytical framework encompassing the innovation environment, innovation actors, and innovation networks. This framework systematically reveals the coupling mechanisms between endogenous and exogenous evolutionary drivers and their impact on collaboration performance. It overcomes the limitations of conventional single-dimensional approaches and elucidates the complex pathways underlying innovation collaboration.

Methodologically, this research integrates fsQCA and NCA: NCA quantitatively assesses the degree of necessity of antecedent conditions, while fsQCA identifies causal asymmetry and conjunctural sufficiency among innovation factors. Through in-depth analysis of the mechanisms underlying innovation driving patterns, we deepen the understanding of how collaborative entities enhance innovation performance, identify differentiated causal pathways across distinct patterns, and offer theoretical support for context-specific collaboration strategies in complex innovation ecosystems.

At the practical level, our findings provide valuable insights into enhancing innovation collaboration. Enhancing innovation collaboration performance requires moving beyond single-factor models and adopting systematically coordinated strategies that leverage the synergistic effects of multiple dimensions—ultimately generating nonlinear superposition effects and producing “1 + 1 > 2” innovation synergies.

Moreover, multiple pathways can lead to high levels of innovation collaboration performance. Collaborating institutions should make strategic decisions based on systematic assessments of their resource endowments, technological capabilities, and external environments. Through strategic planning, they should develop collaboration models aligned with their core competencies.

At the same time, it is essential to recognize the central role of innovation actors in driving collaboration performance. Institutions should strengthen their core functions as innovation actors, enhance technological innovation capabilities, and accumulate collaborative experience to achieve superior innovation collaboration performance.

5.2. Research Limitations and Future Prospects

This study has several limitations. First, it adopts a static perspective to explore the factors influencing innovation collaboration, without fully incorporating the dynamic, evolutionary aspects of time into the framework for collaboration mechanisms. Future research could conduct an in-depth analysis using time-series data, thereby capturing the dynamic interactive effects of various factors more accurately during the collaboration process.

Second, this study focuses on collaborative patents in the field of Natural Language Processing (NLP), which limits the generalizability of the findings. While innovation collaboration in the NLP field is somewhat representative, future research could incorporate policy data, product data, and other multi-source datasets across various fields for comparative analysis. By integrating data from multiple industries and domains, broader applicability and insights into innovation collaboration models can be gained.

Third, collaborative patent data typically focus only on organizations actively involved in collaboration, neglecting those that are not, which may introduce sample selection bias. Additionally, given the complexity and multidimensionality of the innovation collaboration process, this study made necessary trade-offs in selecting the influencing factors. Based on prior research, we selected key representative factors for analysis, but there is an unavoidable risk of overlooking important factors.

Furthermore, since QCA is based on hypothesized causal directions and focuses on configurational patterns of conditions rather than causal mechanisms, it may have limitations in addressing reverse causality. Future research could use panel data or experimental designs to validate the causal mechanisms underlying these paths, thereby enhancing the reliability of causal inferences.