Multi-feature representation-based graph attention networks for predicting potential supply relationships in a large-scale supply chain network

This study aims to predict potential supply relationships within a large-scale supply chain network. Identifying appropriate suppliers can help companies mitigate disruptions in the supply of materials and finances. Furthermore, it offers the companies potential profits by enabling more effective resource allocation and fostering innovation. While previous studies have adopted machine learning approaches, these methods may not fully capture the complexity of network topology. Graph neural network-based methods have recently gained attention as a promising alternative. However, since graph neural network-based methods mainly rely on fixed aggregation weights, these methods often struggle to capture the complexity of supply relationships between companies. This study proposes multi-feature representation-based graph attention networks, which explore hidden topological relationships between companies by incorporating semantic characteristics such as product and network features. Our findings demonstrate that the proposed method outperforms machine learning-based and state-of-the-art graph neural network-based methods. In addition, ablation studies confirm that the proposed components significantly improve prediction performance.