Unraveling the dynamics and identifying the "superstars" of R&D alliances in IUR collaboration: a two-mode network analysis in China

The study examines whether "superstar" actors exist in China’s industry–university–research (IUR) R&D alliance networks and how these networks form and evolve. Motivated by evidence of positive R&D spillovers, particularly when firms collaborate with universities and public research institutes, the paper focuses on China’s rapidly growing strategic emerging industries. It aims to map collaboration, knowledge, and inter-organizational technology networks from joint patent data, analyze their topology over time, and identify temporally dominant actors and technologies that shape knowledge diffusion and innovation outcomes.

Prior work has used network analysis to study collaboration via co-authorship and joint patents. Patent network analysis reveals diverse network structures but lacks unified definitions due to the multi-level nature of R&D alliances. Three main network types are discussed: (1) collaboration networks (organizational cooperation), (2) knowledge networks (flows and combinations of knowledge elements), and (3) inter-organizational technology networks (affiliation-type, linking organizations with technology elements). Two-mode network approaches capture that organizations simultaneously participate in multiple, functionally distinct networks, offering richer understanding of IUR dynamics. Existing studies often consider country-level UIC networks or analyze inter-organizational and technology networks separately; this paper integrates them at the organizational (micro) level to fill this gap.

Scope: Chinese strategic emerging industries (energy-saving and environmental protection; new energy vehicles; high-end equipment manufacturing; new energy; new materials; new generation information technology; bio-industries). Data: China State Intellectual Property Office (SIPO) database covering global patents; keyword-based retrieval aligned with prior studies; manual front-page screening. Collected 26,704 granted patents (1995–2018); after applying criteria (authorized invention patents; applicants include enterprises/universities/research institutes; at least two patent technology categories; removing records not matching category constraints), 11,763 qualified patents remained (retention 44.05%) involving 849 innovation subjects. Knowledge proxy: 4-digit IPC codes (WIPO). Networks constructed: (1) Collaboration network (organizations and their co-assignee relationships), (2) Knowledge network (4-digit IPC nodes linked via co-application by joint patent assignees), (3) Inter-organizational technology (affiliation) network (bipartite links between organizations and IPCs participating in a patent). Time windows: rolling 5-year windows (1995–1999 through 2013–2017; 19 windows). Tools: Sci2 Tool; web crawler for data acquisition and filtering. Metrics: Whole-network—network scale (nodes/edges), density, central potential/centralization, average shortest path (distance), cohesion index, clustering coefficient. Ego-network—degree centrality, betweenness centrality, structural holes (effective size, efficiency, constraint/limit degree). Visualization and topological analysis were performed for representative windows due to space constraints.

- No permanent, single "superstar" dominates the IUR R&D alliance networks. Dominant actors and technologies vary by period. Lifecycle modes observed: formation (U-R, I-U, U, R, I), growth (I-R, I-U, I-U-R), and mature (I-U-R). - Collaboration network growth: edges rose from 102 (1995–1999) to 5,733 (2013–2017). Density declined from 0.117 to 0.007. Clustering coefficient remained high but trended slightly down (e.g., ~0.898 to ~0.841). Average shortest path fluctuated (e.g., 2.347 to around 3.003). Centralization varied with peaks in 2010–2016 (e.g., 0.379 in 2012–2016). Central hubs/hubs’ emergence increased connectivity and cohesion over time. - Knowledge network expansion: edges increased from 14 (1995–1999) to 3,123 (2013–2017). Density fell from 0.156 to 0.005. Clustering coefficient initially increased then decreased (to 0.364 by 2013–2017). Average shortest path increased markedly in later periods (e.g., ~5.163–5.632). Clear gradation emerged with core knowledge elements centrally located and non-core elements on the periphery during growth/maturity. - Inter-organizational technology network: edges grew from 8 (1995–1999) to 6,732 (2013–2017). Average degree increased from 1.600 to 10.221; density showed a decrease–increase–decrease pattern (~0.178 to 0.009). Average shortest path first increased then decreased (inflection around 2003–2007/2005–2009). State Grid Corporation of China remained consistently central. Core IPCs H02J3 (AC circuit devices for trunk/distribution networks) and G01R31 (electrical test/fault detection devices) were repeatedly central across periods. - Ego-network rankings (collaboration network): Early 2000s saw Sinopec and leading universities (e.g., Zhejiang University) with top degree centrality; from 2013 onward, State Grid Corporation of China and China Electric Power Research Institute frequently occupied top positions. Structural holes shifted from universities/research institutes (early 2000s) to research institutes (2005–2010), then enterprises (2011–2013), and from 2014–2018 commonly State Grid Technology College ranked top on structural hole measures. - Ego-network rankings (knowledge network): Early period leaders included C12N15 (biotech/genetic engineering), A61K35/38 (medical preparations). Later periods featured power/electrical and polymer IPCs (e.g., H02J3, G01R31, C08F22). Structural hole positions evolved accordingly (e.g., H02J15, G06Q30/G06N99, G01R21 featured later). - Temporal evolution: Networks densified and integrated over time with more actors and technologies participating. Policy milestones (e.g., strategic emerging industries in 2009, the “2011 plan” collaborative innovation centers) coincided with rising university and enterprise engagement and a transition toward the mature I-U-R mode.

The findings address the central question of whether enduring "superstars" exist by showing that dominance is period-specific and contingent on policy, technological focus, and network growth. While certain actors (notably the State Grid Corporation of China) and technologies (H02J3, G01R31) recur as central, the system exhibits rotating leadership across organizations and knowledge domains. The networks’ structural evolution—from sparse, fragmented clusters to more integrated systems with identifiable hubs—facilitates faster knowledge diffusion and broader recombination. The identified lifecycle (formation, growth, mature) elucidates how collaboration modes shift from bilateral (U–R, I–U) toward a triadic, interdependent I–U–R model. These dynamics underscore the significance of bridging roles (structural holes) and centrality for mobilizing knowledge flows, enhancing innovation performance, and aligning with national policies that stimulate IUR collaboration and the transformation of research outcomes.

This study advances micro-level understanding of IUR collaborations by integrating collaboration, knowledge, and inter-organizational technology networks via two-mode analysis of Chinese joint patents. It maps temporal structural changes, identifies time-bounded "superstars" among organizations and IPC technologies, and characterizes a three-stage evolution culminating in an I–U–R mode. Practically, it highlights the importance of cross-sector integration, sustained firm–university–institute partnerships, and policy support to convert knowledge into impactful technologies. Future research should extend to multi-modal settings, broader patent sources, and benchmark (simulated) comparisons to sharpen causal inferences about network evolution and performance.

Data are limited to SIPO-derived IUR patent records; incorporating WIPO, USPTO, and EPO could improve coverage and generalizability. The analysis employs a two-mode framework but real-world innovation systems are multi-modal (e.g., regional, organizational internal/external knowledge contexts). Additional network indicators (e.g., diameter, efficiency, block models, connectivity, small-world properties) were not fully explored. Distinguishing empirical structures from randomness would benefit from benchmark/simulated networks controlling nodes/edges. These extensions could refine understanding of network lifecycle dynamics and the conditions under which particular actors or technologies achieve centrality.