Innovation Networks’ Hidden Structure Achieves Clarity with Three Strategic Domains

The intricate patterns of collaboration driving innovation remain poorly understood, particularly the underlying structure that shapes inventive activity. Lorenzo Emer, Anna Gallo, and Mattia Marzi, alongside colleagues from the Scuola Superiore Sant’Anna, IMT School for Advanced Studies, and DTU Technical University of Denmark, investigate this problem using extensive patent data across the fields of artificial intelligence, biotechnology, and semiconductors. Their research examines both co-inventorship and co-ownership networks, revealing distinct differences in how inventors and organisations connect. This work is significant because it demonstrates that standard methods for analysing network structure may overlook crucial hierarchical elements, and that understanding these structures is directly linked to the distribution of innovative impact. The team’s findings highlight a concentration of influence within specific clusters, suggesting a non-uniform landscape of inventive contribution.

Inventive Networks in AI, Biotech, Semiconductors

Innovation arises from intricate collaboration patterns between inventors, firms, and institutions, yet the underlying mesoscopic structure governing inventive activity remains largely unexplored. This study addresses this gap by leveraging patent data to analyse both co-inventorship and co-ownership networks within three key strategic domains: artificial intelligence, biotechnology, and semiconductors. Researchers characterised the mesoscale structure of each domain by comparing modularity maximisation, a standard method, with a more refined approach based on minimising the Bayesian Information Criterion, utilising the Stochastic Block Model and its degree-corrected variant. This comparative analysis provides new insights into how inventive activity self-organises across different sectors.

The work reveals that inventor networks are consistently denser and more clustered than those of organisations, a finding consistent with the presence of small, recurring teams operating within larger institutional frameworks. Conversely, organisation networks exhibit neater, hierarchical structures with a limited number of bridging firms coordinating peripheral actors. These distinct network topologies suggest differing coordination logics at play within each level of innovation activity, highlighting how knowledge is created and disseminated. The study employed data from the ORBIS IP patent database, constructing networks from the top 500 most influential actors in each sector, ranked by forward citations between 2020 and 2024.

Further analysis demonstrates a strong connection between these discovered meso-structures and innovation output. Lorenz curves of forward citations consistently show a significant inequality in technological influence, with a concentration of citations within a few key clusters across both inventor and organisation networks. This concentration is particularly pronounced within inventor networks, suggesting that a small number of collaborative groups drive a disproportionate amount of impactful innovation. The results indicate that traditional modularity-based methods may not fully capture the complexities of collaboration and its impact on inventive influence.

Scientists prove that the presence of local hierarchies necessitates more sophisticated tools based on Bayesian inference to accurately model the spread of inventive impact across technological domains. By applying and systematically comparing modularity maximisation with Bayesian Information Criterion minimisation within the Stochastic Block Model framework, the research establishes a more nuanced understanding of innovation network organisation. This work opens new avenues for investigating the interplay between network structure, collaboration patterns, and the generation of technological advancements in critical sectors like artificial intelligence, biotechnology, and semiconductors.

Innovation Network Structure in Key Technologies

The study investigates the mesoscopic structure of innovation networks within artificial intelligence, biotechnology, and semiconductors, employing patent data from 2020 to 2024. Researchers constructed both co-inventorship and co-ownership networks, focusing on the top 500 most influential actors in each sector as determined by forward citation counts, a standard measure of technological impact. This selection criteria ensured a comparable and interpretable dataset across the three domains, capturing the densest and most influential collaborative segments. Network theory provided the framework for representing these systems, with nodes representing actors and edges denoting collaborative links between them.

Scientists pioneered a comparative methodology to characterise the mesoscale structure of each domain, utilising two distinct approaches to identify clusters. A standard modularity maximisation technique served as a baseline, while a more refined method based on the minimization of the Bayesian Information Criterion within the Stochastic Block Model and its degree-corrected variant was also implemented. This innovative approach allowed for a detailed examination of how collaborations shape the dissemination of inventive impact across technological landscapes, addressing limitations of simpler modularity-based methods. The team specifically aimed to capture local hierarchies within the networks, requiring tools grounded in Bayesian inference.

Experiments employed complementary cumulative degree distributions to analyse local network connectivity, revealing key differences between inventor and organisation networks. Results demonstrate that organisation networks consistently exhibit heavier tails, indicating greater heterogeneity among actors compared to inventor networks. This suggests a more uneven distribution of influence and connectivity within the organisational landscape of innovation. The study further reveals that inventor networks are denser and more clustered, indicative of small, recurrent teams operating within larger institutional structures.

Conversely, organisation networks display neater, hierarchical structures with fewer bridging firms coordinating peripheral actors. Lorenz curves of forward citations consistently demonstrate a pervasive inequality in influence, with a concentration of citations within a small number of discovered clusters across both inventor and organisation networks. This finding highlights the crucial role of specific clusters in driving technological advancement and underscores the importance of understanding the underlying network structures that facilitate knowledge diffusion and innovation output.

Inventor and Organisation Network Structures Compared

Scientists have revealed significant insights into the mesoscopic structure of innovation networks across artificial intelligence, biotechnology, and semiconductors. The research team analysed patent data from 2020-2024, constructing both co-inventorship and co-ownership networks using the top 500 most influential actors in each sector, ranked by forward citations, a standard measure of technological impact. This approach allowed for a detailed examination of how collaboration patterns self-organise within these crucial technological domains. Experiments demonstrated that inventor networks consistently exhibit higher density, assortativity, and clustering compared to organisation networks.

Specifically, inventor networks display tighter, recurrent collaboration circles, while organisation networks reveal a sparser, locally tree-like topology indicative of hierarchical structures. Data shows that the degree distributions of inventor and organisation networks differ significantly; organisation-level networks consistently exhibit heavier tails, revealing greater heterogeneity among actors. These findings suggest a fundamental difference in how individuals and organisations engage in collaborative innovation. Further analysis employed both modularity maximisation and Bayesian Information Criterion (BIC) minimisation within the Stochastic Block Model framework to detect mesoscale structures.

Results demonstrate that the baseline modularity-based method struggles to fully capture the hierarchical patterns present in the data. The application of BIC minimisation, however, revealed additional hierarchical and role-based patterns not apparent through modularity alone, highlighting the importance of statistically-grounded inference methods. Lorenz curves of forward citations consistently show a pervasive inequality in technological influence across all sectors and methods. Both inventor and organisation networks demonstrate a high concentration of citations within a few discovered clusters, indicating that influence is not evenly distributed. Measurements confirm that a small number of clusters consistently receive a disproportionately large share of citations, demonstrating the presence of key hubs of innovation within these networks. This work delivers a refined understanding of innovation network architecture and its connection to the dissemination of inventive impact.

Networks Shape Knowledge Creation and Diffusion

This research demonstrates that the structure of innovation networks significantly influences the creation, coordination, and diffusion of technological knowledge within the global patent system. Analysis of co-inventorship and co-ownership networks in artificial intelligence, biotechnology, and semiconductors reveals a consistent pattern across sectors: cohesive groups of inventors connect across organizational and technological boundaries, while organizations exhibit sparser, more hierarchical structures. These findings suggest a dual pattern where collaborative inventor teams drive knowledge creation at a microscopic level, and organizational hierarchies coordinate and amplify this knowledge. The study further establishes a link between these mesoscale network structures and innovation output, specifically demonstrating a concentration of forward citations within a few key clusters.

Importantly, the authors found that standard modularity-based methods may underestimate the importance of local hierarchies in capturing the full impact of collaboration, advocating for more refined Bayesian inference techniques. The analysis focused on the 500 most cited actors between 2020 and 2024, acknowledging a limitation in excluding emerging innovators and a static treatment of network composition. Future research will address these limitations by extending the analysis longitudinally to observe the evolution of these structures and exploring the impact of external factors like policy interventions. Additionally, incorporating citation networks alongside co-inventorship and co-ownership data promises a more comprehensive understanding of knowledge flow. Ultimately, this work highlights the crucial role of understanding the mesoscale architecture of collaboration networks in explaining the distribution of inventive impact across technological domains.