Beyond Complements and Substitutes: A Graph Neural Network Approach for Collaborative Retail Sales Forecasting

Relation-empowered retail management has gained increasing attention. Although evidence on enhanced retail sales forecasting (RSF) for a focal product by leveraging information on related products has been acknowledged, prior studies suffer from high risk of either erroneously introducing irrelevant relations or missing informative ones, as well as incompetence to simultaneously tackle multifaceted and complex product relations. Beyond the well-known relations of complements and substitutes, on the ground of the cross-category choice dependence theory, we discern product relations along both relation type and temporal dimensions: positive and negative relations (from the relation type perspective, with the characteristics of indirectness and asymmetry) as well as asynchronous and dynamic relations (from the temporal perspective). Revolving around how to identify inherent or recurring related products precisely and comprehensively, how to simultaneously leverage indirect, asymmetric, positive, and negative product relations, and how to leverage asynchronous and dynamic product relations, we propose a graph neural network-based method named CL4RSF with a novel data-driven product relation identification strategy and capability of incorporating all above-mentioned product relations. In addition, to further adapt to the RSF context, we design an end-to-end deep-learning architecture equipped with capabilities of multistep forecasting and multisource information fusion. Empirical evaluation on two real-world retail data sets demonstrates the superior forecasting performance of our proposed end-to-end method over state-of-the-art benchmarks and verifies the utility of key designed components in CL4RSF as well as that of leveraging diverse product relations. Further explanatory analyses render insights into cross-category effects and various inferred product relations.