The hidden structure of innovation networks

TL;DR

The paper investigates the mesoscale architecture of innovation networks in AI, biotechnology, and semiconductors using top-500 patent actors. It contrasts modularity-based community detection with Bayesian information criterion (BIC) minimization under Stochastic Block Models (SBM) and degree-corrected SBM (dcSBM) to uncover hierarchical and role-based structures. The results show cohesive inventor teams embedded in sparse, hierarchical organization backbones, with significant inequality in technological impact concentrated in a small number of clusters. These findings highlight the importance of mesoscale inference for understanding invention diffusion and suggest Bayesian approaches capture organizational features that modularity misses, with implications for policy and R&D strategy.

Abstract

Innovation emerges from complex collaboration patterns - among inventors, firms, or institutions. However, not much is known about the overall mesoscopic structure around which inventive activity self-organizes. Here, we tackle this problem by employing patent data to analyze both individual (\emph{co-inventorship}) and organization (\emph{co-ownership}) networks in three strategic domains (\emph{artificial intelligence}, \emph{biotechnology} and \emph{semiconductors}). We characterize the mesoscale structure (in terms of clusters) of each domain by comparing two alternative methods: a standard baseline - modularity maximization - and one based on the minimization of the Bayesian Information Criterion, within the Stochastic Block Model and its degree-corrected variant. We find that, across sectors, inventor networks are denser and more clustered than organization ones - consistent with the presence of small recurrent teams embedded into broader institutional hierarchies - whereas organization networks have neater hierarchical role-based structures, with few bridging firms coordinating the most peripheral ones. We also find that the discovered meso-structures are connected to innovation output. In particular, Lorenz curves of forward citations show a pervasive inequality in technological influence: across sectors and methods, both inventor (especially) and organization networks consistently show high levels of concentration of citations in a few of the discovered clusters. Our results demonstrate that the baseline modularity-based method may not be capable of fully capturing the way collaborations drive the spreading of inventive impact across technological domains. This is due to the presence of local hierarchies that call for more refined tools based on Bayesian inference.

Figures (11)

• Figure 1: Local connectivity of innovation networks. Complementary cumulative degree distributions of inventor (left) and organization (right) networks in the three, strategic sectors of artificial intelligence, biotechnology and semiconductors (only the top-$500$ actors have been considered). Across all sectors, organization-level networks exhibit heavier tails, a feature revealing the higher heterogeneity of the involved actors.
• Figure 2: Local cohesion of innovation networks. Distribution of the average clustering coefficient for inventor- and organization-level networks in the three, strategic sectors of artificial intelligence, biotechnology and semiconductors (only the top-$500$ actors have been considered). Inventor networks are more clustered than organization networks, a feature reflecting the prevalence of small, cohesive teams; organization networks, on the other hand, are more hierarchical, with few firms establishing many, diverse partnerships while most remain specialized.
• Figure 3: Mesoscale structure of the co-inventorship network in the semiconductor sector. The higher density and average clustering coefficient characterizing inventor networks reflect into a pronounced modular structure that points out field-specific specializations leading to project-based collaborations: this, in turn, suggests that knowledge recombines through phases of concentrated teamwork within communities.
• Figure 4: Diversity of the inventors' modules across sectors and models. Each point represents a cluster of the top-$500$ inventor networks in the three, strategic sectors of artificial intelligence (top row), biotechnology (middle row) and semiconductors (bottom row), according to modularity maximization (left column), BIC minimization instantiated with the SBM (middle column) and BIC minimization instantiated with the dcSBM (right column). Ownership diversity (i.e. the number of distinct patent owners), distinguishing single-company from multi-company clusters, is reported on the $x$-axis; technological diversity (i.e. the Shannon entropy of IPC codes), distinguishing specialized from generalist clusters, is reported on the $y$-axis. Dashed lines mark the relevant thresholds: Shannon entropy is split at its median value; organizational diversity is split at $D_{\mathrm{own}}=1$; the bottom-right boxes report the percentage of partitions falling into each quadrant. Across all sectors, modules of inventors are predominantly inter-company and generalist, a feature indicating that inventive collaborations typically involve multiple organizations. To be noticed that modularity maximization leads to individuate broader, more diverse clusters than BIC minimization.
• Figure 5: Mesoscale structure of the co-ownership network in the semiconductor sector. Differently from inventor networks, organization networks are much less modular, appearing as combinations of core-periphery structures: while few firms maintain many cross-institutional partnerships, most actors remain specialized or, at least, regionally confined; this, in turn, suggests that knowledge flows across communities in a hierarchical fashion.