An interpretable product demand forecasting framework incorporating domain knowledge: a case study of substitution effects in community group-buying

With the rapid rise of social e-commerce, an increasing number of consumers now purchase goods through social community networks. However, substitution effects arising from limited inventories and stockouts can seriously impair demand forecasting accuracy. To address this challenge, we propose XAI-Sub, an interpretable demand forecasting framework that systematically incorporates domain knowledge via text mining and biclustering to improve both transparency and predictive performance. By aggregating sparse sales records into substitution-aware clusters and estimating scenario-specific substitution matrices using a novel alternating minimisation algorithm, XAI-Sub explicitly quantifies substitutive relationships while preserving interpretability. Validated on data from a large community group-buying (CGB) platform, our framework achieves a relative improvement of 38% in forecasting accuracy compared with conventional methods. The approach introduces three key advancements in explainable artificial intelligence (XAI): (1) domain-knowledge anchoring: text-derived semantic features anchor substitution patterns within business logic, facilitating human-AI alignment; (2) scenario-driven interpretability: biclustering decomposes demand dynamics into actionable substitution typologies; and (3) causal pathway visualisation: the substitution matrix serves as an interpretable interface, delineating demand redistribution pathways across clusters. This work demonstrates how formalising domain knowledge can bridge the ‘explanation gap’ in complex demand systems and provides CGB enterprises with practical tools to audit and exploit substitution mechanisms.