Classification of Companies using Graph Neural Networks

Abstract

Classification of companies into GICS categories can be addressed using Graph Neural Networks (GNN), by utilizing the different types of relationship between companies such as customer, supplier, partner, competitor, and investor. We use the Relato business graph data and compare the performances of several GNNs and a large language model like BERT that is trained only on the descriptions of the companies. Our goal is company classification into its corresponding category within the four tiers of the GICS hierarchy. Several architectures of GNNs are explored such as GCN, GraphSAGE and GAT, but also RGCN and RGAT that consider the edge type, or relationship between the companies. The main purpose is to reveal what kind of relationship between the companies is most valuable when determining the category of the company. The findings indicate that Graph Neural Networks (GNNs) enhance both classification performance and the understanding of collaboration patterns among companies, providing valuable insights for determining the industry in which these companies operate. This contrasts with the classification based solely on company descriptions using BERT.