INTEGRATING HIGH-UTILITY AND PERIODIC PATTERNS WITH DEEP LEARNING FOR NEXT BASKET RECOMMENDATION

Abstract

Next Basket Recommendation (NBR) predicts a user’s future purchases by analyzing historical transaction sequences. Current deep learning approaches often struggle with variable sequence lengths, information-dense baskets, and the integration of explicit signals like item utility and purchase periodicity. This paper proposes HybridSPMF, an “SPM-First” framework that neuralizes classical sequential pattern mining (SPM) concepts into a learnable architecture. The framework features a Hierarchical Memory Pool to handle histories ranging from 9 to 91 baskets and Adaptive Basket Compression to process dense transactions. Experimental results across three retail datasets demonstrate that HybridSPMF achieves superior performance. Notably, on the Ta-Feng dataset, the model achieves a 56.6% improvement in UtilityRecall@20 and a 24.1% increase in NDCG@20. These findings validate that neuralizing mining patterns enhances both recommendation accuracy and business-centric value.