Technological Breakthrough Tendency in Patent Networks Under Open Innovation: Evidence from Autonomous Driving Patents

Abstract

Firms can gain a competitive advantage through a strategic patent portfolio, wherein patents elucidate technological advancements and establish legal barriers that keep competitors out. However, patents do not provide a perpetual monopoly within the prevailing open innovation paradigm, which means that firms should keep up with innovation input and patent applications to preserve their market dominance. Fostering technological breakthroughs in the patent network thus becomes a critical issue. Anchored in the theoretical views of open innovation, this study conducts an empirical analysis of patent data to examine how patent network structural features influence the technologies’ breakthrough tendency in the field of autonomous driving (AD). The findings indicate that centrality metrics such as degree centrality, harmonic centrality, and betweenness centrality within AD patent networks exert significant influence on technological breakthrough tendency, and the patent family size plays a moderating role in these relationships. Moreover, this research advances theoretical insights for patent strategy formulation in emerging firms of AD, with broader implications for other technology-intensive sectors.

1. Introduction

Patents on emergent technologies constitute a pivotal source of sustainable competitive advantage. By securing patent grants, firms erect innovation barriers rooted in exclusive legal rights [1], thereby safeguarding their proprietary products or services while legally precluding rivals from utilizing identical or substantially similar technologies. Nevertheless, such innovation barriers are temporally bounded in the face of discontinuous technological shifts. Autonomous driving (AD) epitomizes a typical technological breakthrough in the automotive industry. Incumbent firms perpetuate R&D expenditures and patent portfolios to preserve market dominance, seeking to forestall latecomer entrants from appropriating and leveraging the technological breakthroughs. Consequently, a rigorous elucidation of the evolutionary mechanisms underpinning technological breakthroughs at the patent level yields profound managerial implications for firms seeking to capture technological opportunities and optimize their innovation productivity.

The existing research has investigated the antecedents of technological breakthroughs from lenses including stakeholder collaboration strategies [2], absorptive capacity [3], and information resource orchestration [4]. These studies largely concentrate on the organizational level to explore the mechanisms underpinning technological progression, identifying positive effects but overlooking the influence of interfirm network linkages on technological breakthroughs. In practice, firms frequently encounter formidable obstacles in realizing technological breakthroughs, which suggests that latent impediments remain inadequately explicated. Although firm-level analyses can inform subsequent innovation planning, firms often fail to demarcate precise constraints within the patent landscape of emerging technologies, and avoiding these barriers may facilitate a seamless innovation trajectory.

Against the backdrop of the industrial dynamics and theoretical frontiers, this paper poses the following research question: How do patent networks under open innovation influence the breakthrough process in a patented technology? Building upon existing studies on discrete emerging technologies [5,6], this paper approaches this question by focusing on the tendency for technological breakthroughs of AD—specifically, the likelihood of attaining breakthroughs within individual AD patents. Furthermore, open innovation theory serves as the analytical framework for addressing this research inquiry. Given the rapid advancement of AD, firms increasingly rely on external resources to accelerate innovation, yet this reliance entails a dual-edged effect. On the one hand, integrating external resources augments organizational capabilities; conversely, excessive reliance on the outside may engender resource dependency, thereby exposing risks such as supply chain vulnerabilities or lock-in.

To examine the determinants shaping technological breakthroughs, as outlined in the research question, this study selects the AD domain as the empirical context and employs the AD-related patent dataset. The automotive industry has undergone continuous technological evolution since its inception, with no signs of cessation. Influenced by factors such as legal frameworks, industrial policies, and societal transitions, the automobile sector has transitioned from internal combustion powertrains to electrified mobility. The contemporary trajectory is converging toward AD. This progression underscores AD as a paradigmatic breakthrough in the automotive industry, making it an ideal empirical setting with broad practical implications.

Empirical findings reveal that, within the patent networks reflecting the AD open innovation environment, the centrality characteristic of patent entity exerts a negative influence on the propensity for technological breakthroughs. Conversely, the expansion of the patent family exerts a positive moderating effect on the influence relationships between network centrality and technological breakthrough tendency.

2. Literature Review

2.1. Theoretical Evolution of Technological Breakthroughs

From an industrial development perspective, technological breakthroughs fundamentally represent a multifaceted evolutionary trajectory, while within market ecosystems the process signals the gradual obsolescence of incumbent technological regimes. Unlike incremental or complementary innovations attained through improvisation, technological breakthroughs are characterized by distinct discontinuity [7] and heterogeneity [8]. First, the inherent scarcity makes technological breakthroughs difficult to realize, typically necessitating substantial investment of innovation resources for any likelihood of success. Second, technological breakthroughs carry latent risks: as incumbent industries decline, ongoing incremental R&D projects may be prematurely terminated, and the resulting internal losses across the sector must be offset by diversified development in emerging industries [9]. The transition from technological breakthrough to commercialization represents a critical phase for firms seeking to advance the industry lifecycle; should the commercial outcomes fail to meet market expectations, firms betting heavily on emerging technologies may incur severe reversals.

Previous studies have extensively utilized scientometric methods to construct a “technological space” to depict the evolution of technological breakthroughs, essentially representing the path conditions for the ultimate realization. These conditions include various evolution patterns such as periodic divergence–convergence [10], patent family citation [11], and technological search [12]. Despite the significant differences in the connotations, these patterns collectively highlight several important characteristics of technological breakthroughs: first, they are based on patent information as the carrying foundation, with the development history of technological breakthroughs recorded in patent documents and reflected through data mining methods; second, they contain obvious innovative content, with technological breakthroughs bringing about significant functional innovations compared to existing technologies; third, they develop cyclically over time, with the development process of technological breakthroughs involving the mutual substitution of growth and decline trends.

Technological breakthroughs often exhibit a high degree of novelty upon their emergence in specific domains, yet the new technologies tend to converge with mainstream technologies as they evolve—a gradual process of alignment rather than a radical disruption [13]. Significant divergences persist in the existing literature regarding the precise measurement of technological breakthroughs. From the perspective of knowledge flow, such breakthroughs are conceptualized as processes of knowledge recombination that transcend established knowledge boundaries. Prior technologies exert a positive influence on subsequent technological breakthroughs, particularly through mechanisms such as boundary-spanning knowledge integration [14]. Another line of research posits an inverted U-shaped relationship between existing technologies and the likelihood of breakthroughs, moderated by the breadth of knowledge [15]. This perspective further suggests that the potential for technological breakthroughs in mature technologies may not necessarily be lower than that in emerging ones.

Regarding the critical influencing factors of technological breakthroughs, extant research has revealed the interactive effects of multiple elements. Systematic industrial incentive policies, particularly those fostering the refinement of the industrial value chain, serve as significant drivers of technological advancements [16]. Well-defined policy guidance enhances innovators’ commercial expectations and facilitates the formation of the industrial value chain, which in turn leverages the respective R&D strengths of firms, universities, and research institutions. Industrial innovation policies exert a considerable influence on both exploitative and exploratory innovations during technological breakthroughs, with technological path dependency acting as a moderating factor [17]. Furthermore, policymakers—namely, governmental bodies—affect technological breakthroughs through network embeddedness [18], underscoring the dominant role of governmental policy implementation. In addition, the industrial environment in which the technological breakthrough process occurs is critical, requiring the integration of various components and their interactions within the industrial innovation ecosystem to achieve path configuration for technological breakthroughs [19]. Under the influence of “external technology density”, knowledge spillovers from external technologies promote technological breakthroughs at the regional level [20].

2.2. Innovation Strategy for Technological Breakthroughs

From a corporate perspective, technological breakthroughs constitute a critical component of business strategy. Firms promote technological advancements across multiple organizational levels to secure greater competitive advantages. To achieve such breakthroughs, firms integrate their internal innovation resources with the external innovation ecosystem, thereby forming distinctive innovation development strategies. Given the significant heterogeneity among different innovators in the process of achieving breakthrough innovations, the collaborative innovation network within industries exerts multifaceted impacts on technological breakthroughs, wherein proximity strategies play an indirect mediating role [21]. Varied internal reconfigurations differentially influence firms’ network capabilities and reveal optimal transformation pathways at the level of knowledge leapfrogging [22]. Internally, organizations must focus on constructing and refining core technological systems, incorporating extra-domain knowledge to achieve overarching breakthroughs, or—based on diversified technology portfolios—expanding knowledge frontiers through advancements in technological branches.

At the external innovation ecosystem level, technological breakthroughs are inherently a bidirectional process. Firms must proactively adapt to external conditions, such as policy measures, to enhance their absorptive capacity and accelerate the acquisition of innovative resources. The nested architecture of the innovation ecosystem at the corporate level effectively opens the “black box” of organizational boundaries. This enables system coordination to be passively achieved through network effects and ecological effects, ultimately leading to self-evolution [23]. Meanwhile, governmental and other institutional bodies influence the efficiency of accessing innovative resources through policy formulation, thereby shaping corporate innovation strategies. For instance, public services for scientific and technological information resources exert a significantly positive impact on breakthrough innovation, with strategic aggressiveness serving as a moderating factor [4].

Furthermore, the coupling relationship between the development of an enterprise’s technological innovation and the external innovation ecosystem exerts a more profound influence on its technological breakthroughs. First, the formation of such a coupling is premised on an open innovation environment, wherein firms align their internal knowledge capabilities with external resources through external search strategies. Knowledge power exerts a significant positive impact on technological breakthroughs, with cross-boundary search serving as a mediating factor and absorptive capacity acting as a positive moderator [24]. Second, the sustainability of this coupling relies on technology spillover effects, where pronounced spillovers attract greater stakeholder engagement. The willingness of firms to pursue technological breakthroughs intensifies with increasing spillover effects, while policy incentives also tend to shift accordingly [25]. Third, the coupling mechanism is realized through the establishment of collaborative networks or industry alliances, facilitating the integration of internal and external resources. The strategic evolution of stakeholders influences inter-firm collaborative innovation and the formation of innovation alliances, thereby impacting technological breakthroughs [2].

Within the development of the digital economy, digital transformation represents a significant case of technological breakthroughs [26]. Firms’ digital transformation constitutes a novel strategy enabling organizations to promote technological innovation in other related domains—such as enhancing innovation performance and fulfilling policy objectives through technological breakthroughs [27]. On one hand, digital transformation exerts a substantial influence on firms’ breakthrough innovation, wherein firms’ digital-related capabilities play a critical role. For instance, corporate digital capability positively impacts breakthrough service innovation [28], while absorptive capacity exerts a positive moderating effect [3].

Moreover, digital transformation facilitates firms’ re-innovation following technological innovation failures [29]. On the other hand, digital transformation promotes corporate sustainable development via technological breakthroughs [30]. It also enhances the quality of technological innovation outputs, with digital infrastructure serving as a mediating factor [31]. In the rapidly advancing field of artificial intelligence, corporate development strategies must likewise align with innovation resources, both internal and external to the organization. Breakthroughs in AI technology depend on public R&D institutions and leading firms, with high-quality data exerting a significant impact on applications [32].

2.3. Research Gap and the Focus of This Research

Previous studies have extensively explored issues related to technological breakthroughs, such as the measurement of technological breakthroughs and the driving factors of breakthrough innovation, at the quantitative measurement or empirical testing level. These studies have shaped current frontier trends in the field of digital economy, such as digital transformation [3,29,30] and the data element market [27]. However, technological breakthroughs in the context of digital transformation exhibit significant heterogeneity. In some industries, the enabling effect of digitalization is not obvious, such as in scenarios highly dependent on human labour, like handicrafts and small-scale retail. Existing research has not made distinctions based on this characteristic. In terms of the measurement of technological breakthroughs, existing studies usually focus on the organizational level, while there is a lack of research at the more micro-technical level. Although technological breakthroughs are full of uncertainties, they still show the basic characteristics of technological paths from a historical perspective [33]. Exploring the laws of technological breakthroughs at the technical unit level has important theoretical and practical significance. Moreover, few studies have investigated enterprise breakthrough innovation under the background of the patent system, merely treating patents as proxy variables for innovation output, and failing to fully explain the impact of the patent system on enterprise breakthrough innovation. This paper aims to propose a theoretical framework for technological breakthrough tendency by taking AD as an example, which is influenced by open innovation network factors.

3. Theoretical Framework and Research Hypotheses

3.1. Technological Breakthroughs Based on Open Innovation Networks

The fundamental principle of the patent system is the exchange of disclosure for protection, whereby patent holders obtain legal safeguards for their inventions by publicly disclosing patent information, ultimately enabling firms to gain competitive advantages. However, an open intellectual property policy does not necessarily lead to diminished interests for rights holders; on the contrary, it may initially enhance their benefits, though a diminishing return tends to emerge beyond a certain threshold [34,35]. The process of patent filing and technological breakthroughs progresses in parallel, wherein proprietary rights secured through patent protection sustain the competitive edge derived from such innovations. Under the influence of the open innovation paradigm, inter-firm collaboration has intensified, with the core objective of achieving substantial performance improvements for all stakeholders through the creation of shared value [36].

To better understand technological breakthroughs of firms, this paper adopts the concept of technological breakthrough tendency (TBT), which refers to the possibility of a new innovative breakthrough in a technology based on existing knowledge. At the patent level, technological breakthroughs can be characterized by the International Patent Classification (IPC) codes of the patent. In terms of the number of IPC codes, the more IPC codes there are, the higher the technological diversity of the patent. From the perspective of the content of IPC codes, the technical branches indicated by IPC codes reflect the evolution path of the knowledge base involved in the patent. Changes in IPC portfolio correspond to the evolution of technology [37,38]. Further, under the open innovation network, this paper constructs the theoretical framework shown in Figure 1, which is elaborated in Section 3.2 and Section 3.3.

3.2. The Impact of Open Innovation Network Centrality on TBT

From the perspective of patents themselves, the centrality of a patent within an open innovation network can enhance its value [39], primarily reflected in an increased control over market access. However, highly central innovation nodes, by virtue of their deep embeddedness within existing knowledge networks, tend to resist transformative change, leading to what is known as “path dependency” in technological innovation evolution. Strengthening the centrality of patent nodes in open innovation networks, while enhancing market position and leveraging the exclusive rights of patents, may simultaneously constrain the scope for technological breakthroughs. Metrics of centrality in China’s energy sector innovation network reveal that central positions are predominantly occupied by universities rather than firms [8]. Given that firms are the primary drivers of innovation in practice, those situated at the periphery of the network are relieved from bearing the maintenance costs associated with extensive knowledge flows, enabling them to focus more intently on exploratory innovation beyond the dominant technological paradigms.

Previous research has shown that technologies employing breakthrough innovations are often outlier technologies, and firms assemble knowledge from the alliance network [40]. This characteristic leads to the fact that breakthrough innovations are often not recognized by the market at the beginning of their emergence, and only gain status after the understanding of breakthrough innovations gradually deepens. Technological breakthroughs require the acquisition of knowledge from the outside, generating innovative impetus through the knowledge spillover effect. However, absorbing irrelevant knowledge can have a negative effect [20]. From this negative effect, the absorption of external knowledge by firms is not a case of the more the better; rather, accuracy needs to be improved. Otherwise, excessive knowledge embedding may instead lead to innovation obstacles. This paper focuses on three aspects of the degree centrality (DC), harmonic centrality (HC), and betweenness centrality (BC), which reflect the internalization of the patent node in the open innovation network. Therefore, the following hypotheses are proposed:

Hypothesis 1:

The network centrality of the patent node in an open innovation network negatively affects the TBT.

H1-1: 

DC of the patent node exerts a negative influence on TBT.

H1-2: 

HC of the patent node exerts a negative influence on TBT.

H1-3: 

BC of the patent node exerts a negative influence on TBT.

3.3. The Moderating Effect of Patent Family Size

To enhance the competitive advantage derived from patents, firms—particularly multinational corporations—often pursue international patent portfolios centred around core technologies, leading to an expansion in the scale of the patent family. One contributing factor to this phenomenon lies in the territorial nature of patent rights at the legal level; the protective effect conferred by a patent is limited to the jurisdiction of the granting country or region. Existing research employs patent family data as an indicator to assess the global market expansion of firms [41,42]. However, as patents become increasingly integrated into corporate strategy, the impact of patent family size on the trajectory of technological evolution within firms has grown increasingly evident [11].

The expansion of a patent family essentially reflects a firm’s implementation of an outward-oriented open innovation strategy. Over time, the scale of the patent family becomes internalized as part of the firm’s patent asset value [43,44]. Driven by the motive to consolidate their global market position, firms disclose their technological breakthroughs to achieve market pre-emption. Competitors, by accessing these publicly available patent disclosures, gain insights into technological advances, thereby accelerating their own R&D processes and contributing to collective technological progress within the industry. However, under recent trends in intellectual property policy, many internet companies—such as IBM—have adopted the invention pledge strategy [45], disclosing patent information without pursuing patent protection. This approach reflects a corporate development philosophy focused on building an external patent family rather than asserting absolute control over patent rights, highlighting the role of open sharing as a driver of innovation that surpasses the traditional focus on exclusivity. Based on this reasoning, the following hypothesis is proposed:

Hypothesis 2:

The patent family size of the patent node exerts a positive moderating effect on the relationship between network centrality and technological breakthroughs in an open innovation network.

H2-1: 

The patent family size of the patent node will exert a positive impact on the relationship between DC and TBT.

H2-2: 

The patent family size of the patent node will exert a positive impact on the relationship between HC and TBT.

H2-3: 

The patent family size of the patent node will exert a positive impact on the relationship between BC and TBT.

4. Research Method

4.1. Data Sources

The empirical data utilized in this study were sourced from the Derwent Innovations Index (DII) patent database. DII is a commercially authoritative database that undergoes weekly updates, providing researchers with comprehensive and worldwide invention information in the fields of chemistry, electronics, and electrical engineering, as well as engineering technologies. The data coverage of DII incorporates patent records from the majority of intellectual property offices globally, with documented records dating back to 1963. Focusing on the field of AD, this study first developed a corresponding patent retrieval strategy by referencing existing research [38], covering a time span from 2010 to 2021. As a form of documented enterprise innovation outcomes, patent data offer a holistic overview of various aspects of corporate technological innovation. Aligned with the research questions and objectives of this paper, the study conducted mining and analysis of AD patent data to align them with the quantitative indicators pertinent to the empirical analysis.

During the data-cleaning procedure, this study excludes patents with individual assignees and applies a temporal filtering spanning from 2010 to 2021. Consequently, the paper obtains an analytical sample comprising 33,347 entries, with each patent record representing an individual patent family inclusive of several patents. Moreover, the IPC codes were extracted from each patent within the sample to construct a patent–IPC interaction matrix, which serves to delineate the open innovation network within the AD sector.

4.2. Variable Selection

4.2.1. Dependent Variable

The empirical analysis sets TBT as the dependent variable. In this study, TBT is not a direct observation of technological breakthrough phenomena, but a quantitative measurement obtained through deep learning methods and patent data. The prior research employed the number of new IPC classes acquired by firms within a specific time period [4], suggesting that changes in the IPC of patents reflect technological evolution trends. To capture the latent technological relationships between firms and technological domains, this study employs a recursive graph neural network (RGNN) [38] constructed on a firm–IPC bipartite network. The RGNN method can calculate a technical orientation score for each patent; thus, this study treats the resulting scores as a proxy variable for measuring TBT. The process of RGNN is shown in Figure 2.

Specifically, we define a bipartite graph $G_t=\left(F_t\cup I_t,{Edge}_t\right)$ at time period t, where $F_t$ denotes the set of firms and $I_t$ represents the set of IPC subclasses, and ${Edge}_t$ denotes the set of edges connecting firms and subclasses. An ${\mathrm{e}\mathrm{d}\mathrm{g}\mathrm{e}}_{(\mathrm{f},\mathrm{i})}$ ∈ ${\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e}}_{\mathrm{t}}$ is established if firm f has filed at least one patent in IPC subclass i during period t.

(1)

Embedding Initialization: Each node (firm or IPC) is initialized with a low-dimensional embedding vector. The initial embeddings are randomly generated using Xavier initialization to ensure stable training. For nodes that newly appear in later periods, embeddings are initialized in the same manner.

(2)

Representation Learning: The RGNN model learns node representations through a multi-layer message-passing mechanism. At each layer l, the embedding of node v is updated by aggregating information from its neighbouring nodes as follows:

$${\mathrm{e}\mathrm{m}\mathrm{b}}_{\mathrm{v}}^{(\mathrm{l}+1)}=\mathsf{\sigma }\left({\mathrm{W}}^{\mathrm{l}}·\left({\mathrm{e}\mathrm{m}\mathrm{b}}_{\mathrm{v}}^{\mathrm{l}}+\sum _{\mathrm{u}\in \mathrm{N}\left(\mathrm{v}\right)}{\displaystyle \frac{1}{\sqrt{\left|\mathrm{N}\left(\mathrm{v}\right)\right|·\left|\mathrm{N}\left(\mathrm{u}\right)\right|}}}{\mathrm{e}\mathrm{m}\mathrm{b}}_{\mathrm{u}}^{\mathrm{l}}\right)\right)$$

where ${emb}_v^l$ denotes the embedding of node v at layer l, N(v) is the set of neighbouring nodes, $W^l$ is a trainable weight matrix, and σ represents the adopted nonlinear activation function, namely the LeakyReLu.

(3)

Recursive Temporal Updating: To reflect the dynamic evolution of technological activities, the model is trained recursively across multiple time periods. The embeddings learned in period t are used as the initial embeddings for period t + 1. This design allows the model to accumulate historical information and capture temporal continuity in firm–technology relationships. This propagation mechanism enables each node to incorporate both direct and higher-order structural information from the network.

(4)

Latent Preference Score (LPS): After training, the latent preference score between a firm f and an IPC class i is computed as the inner product of their embeddings:

$${\mathrm{L}\mathrm{P}\mathrm{S}}_{(\mathrm{f},\mathrm{i})}={{\mathrm{e}\mathrm{m}\mathrm{b}}_{\mathrm{f}}^{{\mathcal{S}}_3}}^{\mathrm{\top }}{\mathrm{e}\mathrm{m}\mathrm{b}}_{\mathrm{i}}^{{\mathcal{S}}_3}$$

This score reflects the relative affinity between a firm and a technological domain, capturing latent technological orientation inferred from the network structure. Higher LPS values indicate that the patent has a stronger propensity toward specific technological domains.

(5)

Model Training and Convergence: The model is trained using a pairwise ranking objective based on observed firm–IPC interactions. For each firm, observed IPC classes are treated as positive samples, while unobserved IPC classes are sampled as negative instances. The optimization is performed using stochastic gradient descent. Training proceeds until convergence, which is the defined by stabilization of the loss function or reaching a predefined maximum number of epochs. In practice, the loss function decreases steadily and converges within a limited number of iterations, indicating stable embedding representations. This study adopts the RGNN method and transforms the application into an individual patent of an autonomous vehicle, which assesses each patent’s TBT through LPS.

Moreover, this study applies term frequency–inverse document frequency (TF-IDF) weighting to obtain a more accurate result. TF–IDF weighting is applied solely at the aggregation stage to mitigate the dominance of highly frequent IPC classes. After training the RGNN model, we obtain a latent preference score ${LPS}_{fi}$ for each firm f and IPC subgroup i, thus capturing the firm’s latent technological inclination toward the corresponding IPC category. Since the RGNN model operates at the firm–IPC level, an additional aggregation step is required to derive preference measures at the patent level.

Each patent p is associated with a set of IPC classes $I_p=\left\{i_1,i_2,\dots ,i_{\left|I_p\right|}\right\}$. For patents with multiple assignees, the complete set of IPC classes is assigned to each applicant firm, following standard practice in patent-based innovation studies. Accordingly, patent-level preference scores are constructed separately for each firm–patent pair.

To mitigate the dominance of highly frequent and technologically broad IPC classes, we aggregate firm–IPC latent preference scores using TF–IDF-style weights. Specifically, the inverse document frequency (IDF) of an IPC class i is defined as follows:

$${\mathrm{I}\mathrm{D}\mathrm{F}}_{\mathrm{i}}=\mathrm{l}\mathrm{o}\mathrm{g}\left(1+{\displaystyle \frac{\mathrm{N}}{1+{\mathrm{d}\mathrm{f}}_{\mathrm{i}}}}\right),$$

where N denotes the total number of patents in the corpus and ${df}_i$ represents the number of patents assigned to IPC class i. This formulation assigns lower weights to ubiquitous IPC classes and higher weights to more technology-specific ones.

Given the absence of repeated IPC assignments within a single patent, the term frequency component is treated as uniform across IPCs within the same patent. The patent-level preference score of firm f for patent p is then computed as a weighted average of firm–IPC latent preference scores, representing TBT:

$${\mathrm{T}\mathrm{B}\mathrm{T}}_{\mathrm{f}\mathrm{p}}={\displaystyle \frac{{\sum }_{\mathrm{i}\in {\mathrm{I}}_{\mathrm{p}}}{\mathrm{I}\mathrm{D}\mathrm{F}}_{\mathrm{i}}·{\mathrm{L}\mathrm{P}\mathrm{S}}_{\mathrm{f}\mathrm{i}}}{{\sum }_{\mathrm{i}\in {\mathrm{I}}_{\mathrm{p}}}{\mathrm{I}\mathrm{D}\mathrm{F}}_{\mathrm{i}}}}$$

This post hoc aggregation procedure allows us to translate firm–IPC latent representations learned by the RGNN model into firm–patent-level preference measures while preserving the relative technological specificity of IPC classes. We emphasize that the resulting patent-level preference scores are constructed indicators rather than direct model outputs, and they are primarily intended for comparative and analytical purposes.

4.2.2. Independent Variable

Building upon the theoretical postulates and existing studies [8,21,39], this paper empirically employs centrality measures of open innovation networks as independent variables, specifically including DC, HC, and BC, as shown in

Table 1. Independent variables.

Variables	Definition	Exploration
Degree Centrality (DC)	The number of connections that a node establishes with other nodes in a network	This study employs this indicator to quantify the number of IPCs associated with patent nodes within the patent–IPC network, thereby characterizing the technological breadth of individual patents
Harmonic Centrality (HC)	The reciprocal of the harmonic means of the shortest-path distances from a node to all other nodes in a network	This study employs this indicator to quantify the similarity degree between a patent node and other nodes within the patent–IPC network, thereby delineating the specific position of individual patents within the network
Betweenness Centrality (BC)	The probability that a node lies on the shortest path between two other nodes in a network	This paper employs this indicator to measure the bridging role of patent nodes in the patent–IPC network, thereby depicting the technological control ability of individual patents

. The centrality measurements are calculated based on the unprojected patent–IPC bipartite network. This bipartite network is an undirected, unweighted network that primarily characterizes the technical structure of patent nodes, while the centrality indicators further reflect the technical distribution characteristics of patent nodes within the network.

4.2.3. Moderating Variable

Building upon the existing research [11,46,47,48], this paper selects patent family size (Family_size) as a moderating variable in the process of technological breakthrough impacts. A patent family refers to a collection of patent documents sharing a common priority right, characterized by the following features: First, patents within the portfolio are filed and published multiple times across different countries, regions, or international patent organizations, potentially resulting in multiple grant records. Key technologies or high-value inventions promote geographic diffusion through patent families, and this strategy also incurs extra operational costs [49]. Second, the patents in the portfolio are either identical or substantially similar in invention content. Third, the patents within the portfolio exhibit no legal conflicts in terms of validity and are recognized by member parties of relevant international patent conventions.

4.2.4. Control Variable

In terms of control variables, certain inherent characteristics of patents may exert potential influences, necessitating the specification of fixed effects in the model to enhance the robustness of the examined relationships. The following control variables are incorporated into the empirical analysis.

(1)

Number of cited patents in a patent (Cited_patents): This metric refers to the count of prior art patents referenced by a single patent. Patent citations serve as a critical indicator of patent influence, which can be categorized into backward citations and forward citations. Therefore, Cited_patents in this study specifically denote backward citations. This measure reflects the extent to which existing technologies are referenced, bearing a certain correlation with technological breakthroughs, and the citation of prior art also signifies the contribution to innovation and development within the broader technological domain.

(2)

Number of inventors in a patent (Inventor_Num): This indicator refers to the inventor team listed on a patent. As the direct source of patent formation, the knowledge background composition of inventors is closely related to technological breakthroughs. Meanwhile, the information disclosed by inventors in patents also has a potential impact on the performance of patents in open innovation networks.

(3)

Number of present assignees in a patent (Assignee_Num): This indicator quantifies the count of individual entities collectively holding patent rights for a given patent. Similar to inventors, the structure of assignees’ technical expertise influences technological breakthroughs. Furthermore, competitive and collaborative dynamics among assignees contribute additional value to the patent itself.

4.2.5. Instrumental Variable

To address the endogeneity issue of independent variables in the regression models, this study employs the instrumental variable alongside a two-stage regression approach. The instrumental variable Overall_lens is a textual statistic derived from patent documents. Most patent offices usually have formulated restrictions on the length of titles and abstracts in patent documents during the patent examination procedure. For example, the USPTO requires that the title of a patent shall not exceed 500 characters, and the abstract shall contain no more than 150 words [50]. Based on the approximate estimation that one English word equals around 8 characters, this study sets the weights of patent titles and patent abstracts as 0.3 and 0.7, respectively. Furthermore, this study sets the metric Overall_lens through the aggregated text length of both patent titles and abstracts, which is calculated as follows.

$$Overall\_lens=0.3\ast Title\_lens+0.7\ast Abstract\_lens$$

Herein, Title_lens denotes the length of the patent title text, while Abstract_lens refers to the length of the patent abstract text. The weight setting reflects the computational emphasis placed on each component, with the abstract carrying a higher weight due to its richer representation of technological innovations. Both the title and abstract serve as concise summaries of the patent’s core content, encapsulating its value scope. Given that technological innovations documented in patents are primarily conveyed through the title and abstract, these elements facilitate the diffusion of innovation during patent disclosure and form the knowledge foundation for subsequent technological breakthroughs.

Referring to the existing literature [51,52], the selection of instrumental variables should be theoretically reasonable. By controlling for the length of patent documents, the length of patent titles and abstracts does not directly affect TBT, since the increase in text length has a clear upper bound. During the patent application process, the text of patent titles and abstracts is closely associated with patent classification, and the textual keywords therein may directly determine the IPC codes of a patent. Furthermore, individual patents establish connections with other patents through the diffusion of textual information through the public domain, gradually forming patent portfolios centred on specific technical solutions. For innovative entities outside such patent portfolios, if they intend to utilize the relevant technology, they must either obtain patent licencing or develop alternative technologies that circumvent the scope of patent protection. Under this circumstance, innovative entities may achieve technological breakthroughs through R&D activities against the prior patents. Consequently, the patent text length is closely associated with the network topological properties of patent nodes within the open innovation network, making the indicator suitable for instrumental variable analysis.

5. Results

5.1. Descriptive Statistics and Analysis

Based on the mentioned data samples and variable settings, the descriptive statistics of the core empirical variables in this paper are presented in

Table 2. Descriptive Statistics of Core Empirical Variables.

Variables	Means	SD.	Min	Max
TBT	2.541	1.117	0.009	6.383
DC	1.094 × 10−4	9.185 × 10−5	2.410 × 10−5	0.002
HC	10,782.822	1799.287	1.000	14,913
BC	3.962 × 10−5	1.184 × 10−4	0.000	0.004
Family_size	2.034	2.115	1.000	52.00
Cited_patents	10.714	47.542	0.000	1012
Inventor_Num	3.520	2.873	1.000	55.00
Assignee_Num	1.560	1.052	1.000	18.00

.

Table 3. Multicollinearity diagnostic test.

Variables	VIF	1/VIF
DC	2.104	0.475
HC	1.162	0.861
BC	1.538	0.65
Cited_patents	1.062	0.942
Inventor_Num	1.038	0.964
Assignee_Num	2.242	0.446
Family_size	2.389	0.419
Overall_lens	1.082	0.924

presents the collinearity diagnostics for the key variables examined in this study. The results indicate that the Variance Inflation Factor (VIF) values for all variables range between 1 and 10, suggesting that multicollinearity exerts a negligible influence on the outcomes of the empirical model and remains within an acceptable threshold. Since the empirical analysis samples in this paper mainly consist of network data and network centrality indicators, a more stringent collinearity threshold, such as 3 to 5, needs to be adopted. The collinearity diagnosis results show that the VIF values of all variables are below 3, indicating that the empirical analysis sample is still acceptable.

5.2. Baseline Regression Results

This study initially employed a linear regression model to conduct the analysis in JASP 0.95.4, yielding baseline regression results, as presented in

Table 4. Baseline Regression Result.

	M1	M2	M3
DC	−0.279 *** (−46.609)
HC		0.384 *** (74.759)
BC			−0.124 *** (−22.271)
Cited_patents	0.028 *** (5.236)	−0.033 *** (−6.61)	0.006 (1.077)
Inventor_Num	0.035 *** (6.595)	0.023 *** (4.573)	0.028 *** (5.216)
Assignee_Num	−0.067 *** (−11.175)	−0.281 *** (−53.672)	−0.162 *** (−28.852)
N	33,347	33,347	33,347

Note: The coefficients presented in the table are standardized regression coefficients, with t-values shown in parentheses; ***, p < 0.001.

. Degree centrality quantifies the number of direct connections a node possesses within a network. Within the context of the patent–IPC co-occurrence network, it measures the developmental potential of a specific patent node to become a focal point of attention. Conventionally, such degree centrality signifies the importance of a patent node at the relational level and is expected to contribute to the advancement of the technological domain in which the patent resides. However, the results from Model 1 (M1) reveal that degree centrality exerts a negative effect on the propensity for technological breakthroughs, suggesting that an increase in degree centrality does not substantially enhance TBT. Consequently, the hypothesis H1-1 is validated.

As a variant of closeness centrality, HC reflects the relative proximity of a node to all other nodes in a network. Unlike the standard measurement of closeness centrality, HC is calculated based on the reciprocal of distances, where a higher value indicates greater proximity. In patent–IPC co-occurrence networks, this metric embodies the independence and efficiency of technologies. Moreover, the innovation perspective shows that technologies tend to evolve toward simplification rather than complexity. Technologies with simpler knowledge structures are more readily adopted by users in the market, thereby conferring greater competitive advantages. An increase in HC suggests that a technology is more likely to establish knowledge linkages across different technological domains; however, this comes at the cost of reduced adaptability, ultimately constraining technological breakthroughs. In the empirical results of Model 2 (M2), although the baseline regression coefficient of HC is statistically significant, the positive sign does not support the hypothesis H1-2.

BC quantifies the extent to which a node lies on the shortest paths between other nodes within a network. In the context of patent–IPC co-occurrence networks, the indicator reflects the foundational knowledge underpinning technological development—implying that any technological breakthrough necessitates the incorporation of such knowledge. Compared to DC, BC assumes a more significant role as it highlights a node’s function as a “gatekeeper” of information flows. When coupled with patent protection, high BC further reinforces the proprietary nature of technologies, thereby impeding the diffusion of innovation. Empirical results from Model 3 (M3) indicate that BC exerts a negative effect on TBT, implying that an increase in betweenness centrality does not significantly enhance opportunities for technological breakthroughs. Consequently, the hypothesis H1-3 is supported.

Based on the baseline regression, this study employs a stepwise approach by sequentially incorporating control variables, which is to further examine the impact relationships of the independent variables. The results are shown in

Table 5. Regression results with entering variables.

	M4	M5	M6	M7
DC	−0.427 *** (−78.370)		−0.288 *** (−40.718)	−0.420 *** (−66.204)
HC	0.490 *** (99.700)	0.422 *** (82.542)		0.490 *** (99.697)
BC		−0.204 *** (−39.5)	0.016 * (2.416)	−0.013 * (−2.362)
Cited_patents	0.022 *** (4.724)	−0.006 (−1.206)	0.027 *** (5.048)	0.023 *** (4.882)
Inventor_Num	0.033 *** (7.131)	0.024 *** (4.819)	0.035 *** (6.624)	0.033 *** (7.102)
Assignee_Num	−0.116 *** (−22.058)	−0.243 *** (−46.678)	−0.066 *** (−11.089)	−0.116 *** (−22.133)
N	33,347	33,347	33,347	33,347

Note: The coefficients presented in the table are standardized regression coefficients, with t-values shown in parentheses; *, p < 0.05, ***, p < 0.001.

. Given that the baseline regression centres on three independent variables, models (from M4 to M6) introduce each of these three variables as controls in turn, and Model 7 (M7) incorporates all variables, yielding the regression results presented in the table below. Model M4 primarily examines the scenario where DC and HC serve as mutual control variables. It can be observed that both the signs and significance levels of the regression coefficients for these two independent variables remain unchanged, while their magnitudes exhibit a marked decrease and increase, respectively, suggesting independence between the two variables. Similarly, in model M5, the signs and significance levels of the regression coefficients for HC and BC show no change, indicating independence between these variables as well. However, in models M6 and M7, the sign and significance of the regression coefficient for BC undergo notable alterations: its significance diminishes and its sign shifts from negative to positive, implying a correlation between BC and DC. In model M7, the sign and significance of the regression coefficients for DC and HC exhibit no substantial change, further corroborating the preceding analysis.

5.3. Moderating Effect Results

Furthermore, this study introduces patent family size (Family_size) as a moderating variable, which could validate hypothesis H2. Separate regression models, from M8 to M13, were constructed for H2-1, H2-2, and H2-3, with the results summarized in

Table 6. Regression results of moderating effect.

	M8	M9	M10	M11	M12	M13
DC	−0.274 *** (−43.932)	−0.338 *** (−47.361)
HC			0.393 *** (76.82)	0.372 *** (53.82)
BC					−0.113 *** (−19.988)	−0.161 *** (−22.544)
Family_size	−0.019 * (−2.413)	−0.184 *** (−15.192)	−0.165 *** (−22.463)	−0.314 *** (−9.489)	−0.094 *** (−11.802)	−0.137 *** (−15.429)
DC × Family_size		0.202 *** (18.135)
HC × Family_size				0.156 *** (4.630)
BC × Family_size						0.087 *** (10.972)
Cited_patents	0.029 *** (5.402)	0.031 *** (5.858)	−0.020 *** (−4.034)	−0.019 *** (−3.869)	0.012 * (2.191)	0.015 ** (2.661)
Inventor_Num	0.035 *** (6.649)	0.032 *** (6.051)	0.026 *** (5.269)	0.026 *** (5.170)	0.030 *** (5.576)	0.029 *** (5.358)
Assignee_Num	−0.055 *** (−7.034)	−0.015 (0.054)	−0.165 *** (−22.463)	−0.163 *** (−22.143)	−0.097 *** (−12.317)	−0.079 *** (−9.843)
N	33,347	33,347	33,347	33,347	33,347	33,347

Note: The coefficients presented in the table are standardized regression coefficients, with t-values shown in parentheses; *, p < 0.05, **, p < 0.01, ***, p < 0.001.

. An expansion in patent family size signifies an increase in the intrinsic value of the patent [11], encompassing multiple dimensions such as technical, commercial, and legal value. Regression models (M8, M10, and M12) were estimated without incorporating the moderating variable Family_size, while corresponding models (M9, M11, and M13) include this moderating variable. A pairwise comparison of the regression coefficients for the independent variables enables the examination of the moderating variable’s pathway of influence.

As shown in Table 6, the interaction terms in M9, M11, and M13, which interpret the moderating effect of Family_size, are significant and positive simultaneously. Among these, the interactions’ coefficient signs in M9 (b = 0.202), M11 (b = 0.156) and M13 (b = 0.087) are positive. Therefore, Hypothesis 2 and the corresponding sub-hypotheses, H2-1, H2-2 and H2-3, are verified.

However, since the coefficient sign of HC in the baseline regression model is positive, there exist differences in the interpretation of the actual moderating effect of Family_size across the three independent variables. For DC and BC, the main effects are positively moderated by Family_size. The coefficients of DC (b = −0.338) and BC (b = −0.161) bear negative signs, while the coefficients of the corresponding two interaction terms, DC × Family_size (b = 0.202) and BC × Family_size (b = 0.087), are both positive. The coefficient of DC × Family_size indicates that for each one-unit increase in the moderating variable, a positive adjustment of 0.202 units is exerted on the negative effect generated by DC. Correspondingly, the coefficient of BC × Family_size indicates that for each one-unit increase in the moderating variable, a positive adjustment of 0.087 units is exerted on the negative effect generated by BC. Thus, the inclusion of Family_size substantially attenuates the negative influences of DC and BC on TBT. For HC, the main effect is also positively moderated by Family_size. The coefficient of HC × Family_size in M11 (b = 0.156) indicates that the marginal effect of the main effect increases by 0.156 units for each one-unit increase in the moderator variable.

6. Robustness Analysis

6.1. Endogeneity Issue Test

During the preceding empirical testing, while the theoretical hypotheses of this study were largely corroborated, thereby affirming the appropriateness of the empirical analytical approach and the reliability of the sample data, empirical findings across various strata suggested potential endogeneity concerns in the model’s independent variables. To ensure the robustness of the empirical research outcomes, this paper employs the instrumental variable approach along with the two-stage least squares (2SLS) method for further validation, thereby addressing the potential endogeneity issue. The following two tables present the analysis and resolution of the endogeneity problem:

Table 7. The first-stage regression results in the 2SLS.

First-Stage Regression	M14	M15	M16
Overall_lens	4.61 × 10−8 *** (32.12)	0.60 *** (18.97)	1.40 × 10−8 *** (6.73)
Cited_patents	2.55 × 10−7 *** (27.62)	1.66 *** (8.14)	3.43 × 10−7 *** (25.69)
Inventor_Num	8.35 × 10−7 *** (5.50)	7.68 * (2.29)	2.65 × 10−7 (1.21)
Assignee_Num	3.77 × 10−5 *** (87.60)	381.41 *** (40.13)	2.51 × 10−5 *** (40.32)
F test of excluded instruments	F = 1031.64	F = 359.93	F = 45.32
Underidentification test	p = 0.000	p = 0.000	p = 0.000
Weak identification test	F = 1031.64 (>10% maximal IV size 16.38)	F = 359.93 (>10% maximal IV size 16.38)	F = 45.32 (>10% maximal IV size 16.38)
Weak-instrument-robust inference test	p = 0.000	p = 0.000	p = 0.000

Note: The coefficients presented in the table are standardized regression coefficients, with t-values shown in parentheses; *, p < 0.05, ***, p < 0.001.

displays the first-stage regression results, while

Table 8. The second-stage regression results in the 2SLS.

IV (2SLS) Estimation	M17	M18	M19
DC	−6661.19 *** (−15.50)
HC		−5.11 × 10−4 *** (−10.19)
BC			−21,957.77 *** (−6.47)
Cited_patents	1.58 × 10−3 *** (8.93)	7.34 × 10−4 ** (3.34)	7.42 × 10−3 *** (6.00)
Inventor_Num	1.48 × 10−2 *** (6.97)	1.31 × 10−2 *** (4.10)	1.50 × 10−2 ** (2.96)
Assignee_Num	6.86 × 10−2 *** (3.76)	1.21 × 10−2 (0.54)	0.37 *** (4.15)
Endogeneity test	p = 0.000	p = 0.000	p = 0.000

Note: The coefficients presented in the table are standardized regression coefficients, with t-values shown in parentheses; **, p < 0.01, ***, p < 0.001.

shows the second-stage regression results.

As indicated in Table 7, the regression coefficients of Overall_lens in the three models (M14–M16) are statistically significant and positive, suggesting a broadly consistent relationship between the instrumental variable and the independent variable. The F-test of excluded instruments yields F-statistics for all three variables exceeding 10, indicating the validity of the instrument variable. The weak identification test further reveals that the Cragg–Donald Wald F-statistics for all three models surpass the 10% maximal IV size critical value, thus confirming the passage result of the test. These test results collectively demonstrate that the instrument variable is effectively identified in the first-stage regression model, reaffirming the robustness of the 2SLS estimation.

The introduction of instrumental variables serves to isolate the exogenous component from endogenous variables, thereby enhancing the consistency of results of empirical analysis. Since the number of instrumental variables in this study equals that of the endogenous variables, there is no need to conduct an overidentification test for the 2SLS model. The endogeneity test results indicate that all three independent variables exhibit endogeneity. In the regression outcomes of the models (from M17 to M19), the regression coefficients of the three independent variables are all negative and statistically significant, demonstrating the validity of the instrumental variable approach. Furthermore, the negative effects of variables DC and BC are consistent with the baseline regression results, confirming the robustness of the baseline findings through instrumental variable validation. In contrast, variable HC exhibited a positive coefficient in the baseline regression, but its coefficient sign reversed while remaining significant in the 2SLS model. This suggests that HC suffers from severe endogeneity issues. The baseline regression model for HC is biased, whereas the 2SLS approach, by isolating the endogenous component of this variable, reveals the true causal effect, thereby verifying the original hypothesis H1-2.

The results from Table 7 and Table 8 indicate that all three independent variables under investigation exhibit endogeneity issues, while the selected instrumental variable, Overall_lens, demonstrates a statistically significant positive correlation with each of the three independent variables. Furthermore, the instrumental variable successfully passes a series of validation tests. Upon the introduction of the instrumental variable, the endogeneity issue in the baseline regression model has been appropriately addressed, thereby reinforcing the robustness of the research hypotheses.

6.2. Heterogeneity Analysis

The evolution of AD technology constitutes a long-term and multifaceted systemic process characterized by numerous technical branches. To further investigate the robustness of the empirical findings, this section conducts a heterogeneity analysis along two dimensions: temporal intervals and technological categories. The distribution of samples under these classification schemes is displayed in

Table 9. Sample classification of heterogeneity analysis.

	Year_Range	IPC_Section
Sample Classification	- Range 2010–2013 (1): 1360 - Range 2014–2017 (2): 5985 - Range 2018–2021 (3): 26,002	- Section B (1): 15,955 - Section G (2): 14,693 - Section H (3): 1129 - Section A (4): 1008 - Section E (5): 306 - Section F (6): 233 - Section C (7): 23 - Section D (8): 0

. Temporal intervals are divided into three-year ranges (Year_Range) based on patent application dates. Although the IPC system comprises eight main sections (IPC_Section), the patent dataset in this study spans only seven of them. The section codes are replaced by dummy variables from 1 to 8, as shown in Table 9. The numerical codes of IPC categories serve merely for coding convenience in empirical computation.

Policymakers in implementation contexts must carefully consider the timing of policy enforcement and tailor measures specifically for key segments of the industrial chain to ensure desired outcomes. Regarding technological development, both temporal intervals and technical categories serve as critical drivers of technological evolution [37]. On the one hand, technological breakthroughs require sustained long-term accumulation, with core technologies demanding particularly extended R&D periods. On the other hand, advancements vary significantly across different technical branches. To examine heterogeneity in technological breakthrough processes, this study employs an interaction term approach, converting categorical variables into dummy variables. The numerical values assigned to each dummy variable category are indicated within parentheses in Table 9.

The results of the heterogeneity analysis regarding the temporal intervals are presented in

Table 10. Heterogeneity analysis of time range aspect.

	M20	M21	M22
DC	−0.263 *** (−9.383)
HC		0.384 *** (16.478)
BC			−0.095 ** (−3.956)
Year_Range	0.114 *** (14.806)	0.053 * (2.002)	0.082 *** (14.607)
DC × Year_Range	−0.032 (−1.129)
HC × Year_Range		−0.006 (−0.162)
BC × Year_Range			−0.033 (−1.358)
Cited_patents	0.036 *** (6.688)	−0.030 *** (−6.042)	0.011 * (1.957)
Inventor_Num	0.031 *** (5.93)	0.021 *** (4.209)	0.025 *** (4.668)
Assignee_Num	−0.047 *** (−7.821)	−0.274 *** (−51.947)	−0.152 *** (−26.896)
N	33,347	33,347	33,347

Note: The coefficients presented in the table are standardized regression coefficients, with t-values shown in parentheses; *, p < 0.05, **, p < 0.01, ***, p < 0.001.

. As indicated by the interaction terms in the table, the regression coefficients for all three interaction terms are statistically insignificant, suggesting the absence of heterogeneity across the temporal intervals. Furthermore, the regression coefficients for the three independent variables are all significant, and their signs are consistent with those in the baseline regression. These findings demonstrate that the baseline regression results are robust with respect to temporal intervals. Despite the marked differences in the number of patents across the three-year ranges, the overall evolutionary process remains fundamentally consistent. Moreover, this study performed the heterogeneity analysis at the year level (see Supplementary Materials). The supplementary results show that the baseline model remains stable in the short term, but temporal heterogeneity effects appear, mainly presenting negative moderating effects on the impact of TBT in DC and BC, respectively.

The results of the heterogeneity analysis at the IPC section level are presented in

Table 11. Heterogeneity analysis of IPC section aspect.

	M23	M24	M25
DC	−0.218 *** (−19.436)
HC		0.386 *** (41.193)
BC			−0.117 *** (−11.766)
IPC_Section	−0.050 *** (−6.057)	−0.047 * (−2.324)	−0.082 *** (−14.447)
DC × IPC_Section	−0.088 *** (−7.050)
HC × IPC_Section		−0.015 (−0.690)
BC × IPC_Section			−0.010 (−0.976)
Cited_patents	0.030 *** (5.750)	−0.032 *** (−6.446)	0.007 (1.263)
Inventor_Num	0.036 *** (6.884)	0.024 *** (4.720)	0.029 *** (5.385)
Assignee_Num	−0.063 *** (−10.605)	−0.280 *** (−53.347)	−0.161 *** (−28.722)
N	33,347	33,347	33,347

Note: The coefficients presented in the table are standardized regression coefficients, with t-values shown in parentheses; *, p < 0.05, ***, p < 0.001.

. The influence of the three independent variables on TBT remains significant, and the coefficient signs remain consistent, indicating that the analysis results of the baseline regression are robust. However, only the regression coefficient for interaction term DC × IPC_Section is statistically significant, and the other two interaction terms are not significant, revealing the presence of heterogeneity at the IPC section level as well as distinct influence pathways. Specifically, IPC mainly exerts a negative moderating effect on the relationship between DC and TBT. To further verify the heterogeneous impact within IPC, this study also conducted an analysis at the IPC class level (see Supplementary Materials). At this point, the performance of the baseline model remains stable, but the heterogeneity of IPC becomes more pronounced, exerting positive, negative, and positive moderating effects on the impact of TBT in DC, HC, and BC, respectively.

6.3. Alternative Variables Analysis

As the dependent variable of this study is constructed based on deep learning methods and patent data, there is a possibility that this indicator is dependent on specific methods. Therefore, the benchmark model and the moderating effect model also need to adopt the alternative variable approach to test the reliability of the core hypotheses and conclusions of this paper. Referring to existing studies [53,54], this paper selects two different methods, Neural Matrix Factorization (NMF) and Neural Graph Collaborative Filtering (NGCF), to construct new dependent variables, TBT_NMF and TBT_NGCF. The robustness test results can be seen in

Table 12. Robustness test with alternative variables.

	M26	M27	M28	M29	M30	M31
DC	−0.085 *** (−11.522)			−0.380 *** (−54.446)
HC		0.052 *** (6.957)			0.157 *** (21.549)
BC			−0.061 *** (−8.424)			−0.182 *** (−25.793)
Family_size	−0.168 *** (−13.342)	−0.426 *** (−11.825)	−0.143 *** (−15.942)	−0.243 *** (−20.533)	−0.362 *** (−10.361)	−0.180 *** (−20.506)
DC × Family_size	0.066 *** (5.710)			0.250 *** (22.968)
HC × Family_size		0.289 *** (7.908)			0.198 *** (5.576)
BC × Family_size			0.032 *** (4.034)			0.115 *** (14.727)
Cited_patents	0.055 *** (9.920)	0.044 *** (8.053)	0.053 *** (9.597)	0.025 *** (4.846)	−0.019 *** (−3.594)	0.006 (1.203)
Inventor_Num	0.037 *** (6.695)	0.035 *** (6.371)	0.036 *** (6.593)	0.015 ** (2.968)	0.012 * (2.170)	0.012 * (2.303)
Assignee_Num	−0.019 * (−2.275)	−0.053 *** (−6.662)	−0.034 *** (−4.196)	−0.017 * (−2.149)	−0.145 *** (−18.709)	−0.089 *** (−11.192)
N	33,347	33,347	33,347	33,347	33,347	33,347

Note: The coefficients presented in the table are standardized regression coefficients, with t-values shown in parentheses; *, p < 0.05, **, p < 0.01, ***, p < 0.001.

. The first three models (M26 to M28) and the second three models (M29 to M31) simultaneously show the robustness of the baseline model and the moderating model. In the Supplementary Materials, there are more comprehensive results of the alternative variables analysis. The computation implemented by NMF and NGCF, consistent with that of RGNN, is based on the analysis and mining of patent data. The indicator results obtained are used as proxy variables for TBT, rather than a direct observation of technological breakthroughs.

7. Discussion and Conclusions

7.1. Theoretical Discussion

This study undertakes an empirical analysis of technological breakthroughs at the industrial level of AD. In the era of digitization and intelligence, open innovation constitutes a pivotal imperative for firms pursuing technological breakthroughs [55]. Emerging trends in AD suggest that companies are harnessing open innovation to forge collaborative innovation ecosystems, thereby embedding technological breakthroughs into the industrial cluster planning. The empirical results illuminate the evolutionary patterns of TBT, identifying the core factors of network structure. Through empirical assessments of individual patents, this study underscores the pivotal role of patent network centrality in driving technological breakthroughs. The following research conclusions are derived from patents on AD, and can also provide replicable and generalizable theoretical insights for other industrial fields.

First, baseline regression results reveal that nodal centrality within open innovation networks exerts a significant negative impact on TBT. Robustness analyses show that DC and BC have significant effects. This suggests that firms must escape the “capture” of existing technologies when fostering technological breakthroughs, which also explains why peripheral nodes in innovation networks exhibit higher probabilities of generating such breakthroughs. Moreover, it is interesting that there is serious endogeneity in HC. This finding implies that there may exist a bidirectional causal relationship between HC and TBT, or that the observed correlation is affected by other important factors associated with patent networks, which provides corresponding implications for subsequent research.

Second, the analysis of moderating effects indicates that patent family size serves as a significant positive moderator. This result implies that the negative impact of patent node network centrality on TBT changes as the size of the patent family expands. Generally, the larger the patent family, the smaller the hindrance to technological breakthroughs caused by increased network centrality of the patent node, indicating that path dependence within technology can be mitigated through external diffusion. The firm’s technological innovation becomes constrained by path dependency, which may hinder subsequent breakthroughs and create developmental bottlenecks. However, firms with substantial technological accumulation could secure competitive advantages in the short term by expanding their international patent portfolios—thereby increasing patent family size.

Third, robustness analyses confirm that the study’s conclusions remain stable despite potential endogeneity in the independent variables, and the empirical results remain robust as time passes and IPC changes. The hypothesized relationships exhibit no heterogeneity in the long timespan, but show some differences in the short term. In addition, distinct heterogeneity emerges across technology categories, mainly as the main effect of DC. The patents that the technical content aligns with Physical Science (Section G) more closely, will exhibit more pronounced heterogeneity. Firms seeking technological breakthroughs, particularly in traditional sectors such as the automotive industry, should promote the diffusion of digital technologies into non-digital domains to enhance their likelihood of technological breakthroughs.

7.2. Practical Implications

7.2.1. Technical Management for AD

AD is a complex technological field that involves multiple technical branches, relies on big data-driven approaches, and faces extremely high safety regulatory standards. Therefore, in addition to exploring the driving mechanism of TBT from a theoretical perspective, this study can also bring certain substantive insights to the technical management of AD through theoretical discoveries. Currently, the technological development of AD is manifested in specific challenges such as vehicle platooning [56], mobility as a service (MaaS) [57], or trajectory tracking [58]. Moreover, the realization of the core functions of AD also depends on the application of related technologies such as 3D vision applications [59] and trajectory prediction [60]. According to the division of IPC, AD patents are mainly concentrated in Section B and Section G. At the class level, the patents are primarily distributed in the technical categories such as B60W, G05D, B62D, G01C, and B60L. The meaning of the main IPC codes could be seen in the Supplementary Materials. The heterogeneity analysis conducted based on the IPC aspect demonstrates that the technological breakthroughs of AD are closely associated with the technical fields to which the technologies belong. Moreover, there are some technical management suggestions.

First, in response to the issues of weak operational stability and insufficient anti-interference capability of AD vehicles in complex traffic scenarios, R&D plans can be initiated based on technical branches such as G05D and G08G to optimize vehicle trajectory control algorithms, and technical tests can be conducted for multi-scenario adaptation. Second, to address the problems of excessive energy consumption and rough management in the operation of AD vehicles, research and development of on-board electronic control technology can be carried out based on technical branches such as B60W and B60L to reduce the energy consumption per vehicle and expand the applicable scenarios. Third, to tackle the problems of unbalanced transport capacity and low resource utilization in AD dispatching systems, large models and intelligent agents can be introduced based on technical branches, such as G08G and G06Q, to build a dynamic dispatching system, thereby promoting the large-scale commercial application of AD.

7.2.2. Policy Implications for Industrial Planning

The theoretical contributions of this study provide empirical guidance for firms designing innovation strategies and technological competitiveness plans. Moreover, the findings facilitate commercial forecasting for the AD industry and other emerging technology sectors, which also offer a series of managerial implications for the construction and expansion of industrial clusters.

First, industry policymakers of AD should prioritize technological domains represented by outlier nodes in the innovation network, guiding firms across the industrial chain to accelerate technological breakthroughs through industrial planning or policy interventions. Outlier patent nodes refer to specific technologies that are in early-stage development and not yet widely understood by stakeholders within the industrial sector. As empirical evidence accumulates regarding the strategic value of such outlier patents, the associated technologies exhibit higher breakthrough potential relative to more mature or widely adopted alternatives. Industrial policies should channel additional innovation resources toward outlier patents, enabling relevant firms to expedite the realization of value derived from technological innovation.

Second, corporate decision-makers of AD should formulate reasonable expansion strategies for patent families. At the technical level, when core technologies are applied for international patent protection, these patents could occupy larger market shares, but meanwhile conceal the potential risks of technological path dependence. Firms need to comprehensively assess the technological development trends in AD. Once a technology’s competitive advantage begins to diminish in the market, firms should intensify R&D efforts toward next-generation technologies and undertake corresponding patent landscaping. Internally, decision-makers may foster competition among different technological pathways, prioritizing international patent applications for routes demonstrating stronger competitive potential to systematically enlarge their patent family. However, such expansion strategies must evolve in accordance with the maturity of technological trajectories. When the growth rate of technological innovation decelerates, the corresponding patent family expansion strategies should be scaled back accordingly.

Third, technology innovation alliances—particularly those focused on AD—should assume a more central role in orchestrating collaborative innovation, leveraging technological complementarity to drive collective breakthroughs across member organizations. Knowledge exchange among diverse technological domains exerts a profound impact on achieving innovative advancements. Such alliances must effectively integrate stakeholders from various industrial sectors, especially for the automotive industry and the information and communication industry. Industrial organizations could also enhance the dissemination of cutting-edge technological innovations. Firms should intensify external technological collaborations on AD, thereby reinforcing network connectivity of emerging technologies and co-constructing an innovation ecosystem that enables equitable value sharing among participants.

7.3. Limitations and Further Research

This study has several limitations that warrant further investigation in future research. First, the study primarily focuses on the AD industry, and this narrow scope may overlook the impact of industrial heterogeneity on the research conclusions. Although our heterogeneity analysis explores potential influences through the deconstruction of technological subfields, there remains a possibility of omitted variables. Consequently, future studies are planned to extend empirical investigations to a wider range of industries. Second, the empirical analysis in this paper centres on the effect of network centrality on TBTl however, it lacks a systematic explanation for the underlying mechanisms driving this relationship. We also recognize that as a proxy variable, TBT has limitations in measurement, and there is still room for improvement in TBT, which could provide a better understanding of technological breakthroughs. Thus, subsequent research intends to conduct coding analyses of representative AD firms through case studies, with the aim of uncovering the fundamental logic behind firms’ technological breakthroughs.