Consensus Bayesian Network Learning for Market Basket Analysis in Indonesian Retail Transactions

Understanding relationships among products in retail transactions is essential for supporting strategic decisions such as product placement and cross-selling. However, traditional market basket analysis is limited to pairwise associations and cannot capture complex conditional dependencies among multiple items. This study addresses these limitations by applying a Bayesian Network framework to model product relationships probabilistically. Transaction data collected over one month were transformed into a binary dataset of 441 transactions across 31 product categories. The network structure was learned using a stochastic search algorithm with a Bayesian Dirichlet equivalent (BDe) score, enhanced by a multi-seed strategy and edge support aggregation to ensure robustness. Parameters were estimated using maximum likelihood. The results indicate that relationships are both directional and context-dependent, with snack probabilities increasing with instant noodles but decreasing when cigarettes are also present. The LOIO accuracy is 0.6659.