Business partner selection considering supply-chain centralities and causalities

Abstract

ABSTRACT While network centricity-based supplier recommendation models can make predictions with high accuracy, they do not sufficiently encompass supply chain causality, and their interpretability is controversial. We propose adding conditional probabilities from the Bayesian network to a supplier predictive model to improve interpretability while maintaining the performance of the social network approach. We construct a supplier forecasting model for 327,012 corporate transactions in the Northeast region of Japan using corporate attributes, network centrality, and conditional probabilities as features and discuss both performance and interpretability. Random forest, support vector machine, and logistic regression were applied as classifiers, and the outputs were compared. The proposed model exceeded an F1 score of 80%, and we found that conditional probabilities have resulted in the highest significance. By incorporating causal features, we were able to construct an accurate and interpretable model. Our findings have implications for companies’ choice of suppliers and for local governments’ consideration of regional industrial policy from a macro perspective.