Problem

The paper addresses the limited understanding of the overall mesoscopic structure within innovation networks, which are crucial for knowledge creation, recombination, and diffusion. While individual collaborations are studied, the larger, emergent patterns of interaction among inventors, firms, and institutions remain underexplored. The study aims to uncover these hidden structures across various strategic technological domains and understand their relationship with innovation output.

Method

The researchers analyzed patent data from 2020-2024 within Artificial Intelligence, Biotechnology, and Semiconductors, focusing on the top-500 most influential actors. They constructed two types of networks: individual-level (co-inventorship) and organization-level (co-ownership). To detect mesoscale structures, two primary methods were systematically compared: standard modularity maximization (a baseline for community detection) and Bayesian Information Criterion (BIC) minimization within the Stochastic Block Model (SBM) framework and its degree-corrected variant (dcSBM). The detected structures were then characterized using metrics like link density, average clustering coefficient, and internal composition (organizational and technological diversity). The impact of these structures was assessed by analyzing the distribution of forward citations using Lorenz curves and Gini coefficients.

Results

The study yielded three major findings: First, inventor networks were consistently denser, more assortative, and more clustered than organization networks across all domains, suggesting the presence of tightly-knit, recurrent collaboration teams. Conversely, organization networks exhibited sparser, more hierarchical structures. Second, the inference-based SBM framework (BIC minimization) revealed additional hierarchical and role-based patterns that modularity maximization alone failed to capture, indicating its inadequacy for fully understanding complex innovation network organization. Third, differences in mesoscale organization correlated with inventive impact inequality. Inventor networks showed pronounced inequality in patent influence across all sectors, with a small fraction of inventors accounting for a disproportionate share of technological impact. Organizationally, AI exhibited the strongest concentration of citations in a few dominant alliances, while Biotechnology and Semiconductors showed more balanced distributions. The choice of algorithm significantly affected the perceived inequality, with BIC minimization revealing steeper Lorenz curves due to its ability to identify cohesive, high-impact cores.

Implications

The findings demonstrate a dual, nested architecture within innovation ecosystems: cohesive, project-based inventor teams at the micro-level support knowledge creation and exploitation, while sparser, hierarchical organizational structures at the meso-level coordinate and facilitate knowledge diffusion and exploration across broader systems. This multi-scale organization is vital for understanding how technological knowledge is generated and spread. The pervasive inequality in technological impact, especially within inventor networks and AI organizations, suggests cumulative advantage mechanisms are at play. The study also implies that relying solely on modularity-based community detection may obscure crucial organizational features, advocating for more refined, inference-based tools like BIC minimization to accurately model the hidden structure and impact distribution in innovation networks. Future work could extend this analysis longitudinally and to multi-layer network representations.