Innovation and artificial intelligence: a machine learning-based approach to patents

Purpose This study investigates the configurational effects of patent cooperation network and knowledge diversity on the persistent innovation of AI-related firms. It extends social network theory and the resource-based view by revealing how multidimensional network structures and internal knowledge characteristic interact to sustain long-term innovation. Design/methodology/approach The study employs a random forest model to evaluate the relative importance of multi-dimensional collaboration network features, including centrality, structural holes, collaboration depth, and network density. Based on the algorithm-selected key variables, a CART decision tree model incorporates firms' knowledge diversity to uncover configurational patterns driving persistent innovation. The machine learning-based framework enables both variable prioritization and the identification of nonlinear decision pathways that characterize heterogeneous innovation strategies. Findings The analysis identifies eight representative innovation pathways, which are further categorized into three archetypes of AI-related firms: core-bridged, dense-pluralistic, and collaboration-balanced. Core-bridged firms benefit from high network centrality and bridging positions that enhance external knowledge access and long-term innovation capacity. Dense-pluralistic firms leverage cohesive external ties and high internal knowledge heterogeneity to promote sustained exploration. Collaboration-balanced firms, meanwhile, exhibit moderate but well-coordinated connections across multiple domains. These typologies illustrate the equifinal configurations through which diverse firms achieve persistent innovation. Originality/value This study offers a novel, configuration-based understanding of how AI-related firms sustain persistent innovation. By integrating social network theory and the resource-based view, it bridges the structural perspective of external collaboration with the capability-based view of internal knowledge diversity. The analysis highlights that innovation persistence depends not on single factors but on specific combinations of network positions and knowledge configurations. Methodologically, the study advances innovation research by employing an interpretable machine learning framework that combines random forest–based feature selection with decision tree–based pathway extraction, enabling the identification of multiple, equivalent innovation strategies beyond traditional linear assumptions.