[3I5-GS-11] AI and Society:

[3I5-GS-11-05] Transaction Value Estimation in Inter-Firm Networks Using a Hybrid Model Combining Graph Neural Networks and LightGBM

〇Ryoji Sato¹, Ryosuke yano¹, Takayuki Mizuno² (1. Tokio Marine dR Co., Ltd., 2. National Institute of Informatics)

Keywords:Network Science, Supply Chain, GNN

Estimating inter-firm transaction values from limited data is crucial for understanding shock propagation through networks, especially during disasters. Previous studies have employed gravity models, which are commonly used in international trade. In this study, we leverage supervised machine learning to achieve more accurate estimation. Our explanatory variables include firm-level attributes such as sales, industry category, country, and the number of network links. We applied three machine learning methods—LightGBM, multilayer perceptron (MLP), and Graph Convolutional Networks (GCN), a type of graph neural network—and evaluated their predictive performance. LightGBM achieved the highest accuracy, followed by MLP and GCN. This result indicates that incorporating graph structures through GCN does not necessarily enhance prediction accuracy. To further improve performance, we extract GCN-generated node embeddings and incorporate them as additional explanatory variables into LightGBM. This hybrid approach demonstrated higher prediction accuracy compared to using LightGBM alone.