Multi-View Representation Learning for Regional Opportunity Spaces and Industrial Path Development

Abstract

Regional opportunities for industrial development arise when local conditions favor particular economic activities. These conditions depend in part on a region’s existing industry profile and on the economic relationships among its activities. Heterogeneous data sources on mobility, production, and firm ownership provide complementary evidence about this underlying industrial structure. We therefore develop a multi-view representation-learning framework for industries that integrates these different sources into a shared latent representation. By using graph diffusion, the framework captures both direct and indirect connections between industries. Applied to Swedish industries over three decades and three classification versions, the learned representation places industries from different classification periods within a common latent space. To evaluate this operationalization of opportunity, we construct an opportunity-density measure by combining latent similarity with local industry size. Higher opportunity density is consistently associated with industry entry and startup formation and with lower industry exit. Across classification periods, a one-standard-deviation increase in opportunity density is associated with 5.3–6.2 percentage-point higher industry-entry probabilities (118–136% of baseline rates), 8.9–13.8 percentage-point lower industry-exit probabilities (147–209%), and 1.1–1.7 percentage-point higher startup-formation probabilities (131–187%). Opportunity density remains associated with these outcomes in 41 of 42 models after accounting for information contained in the individual data sources. Our machine-learning framework shows how heterogeneous data can be integrated into a common latent structure and used to operationalize regional opportunity spaces, providing a new approach to studying industrial path development.