Conv-based Temporal Sets Prediction for Next-basket Recommendation

Abstract

Next-basket recommendation (NBR) aims to recommend a set of items to a consumer according to past shopping carts. NBR essientially a more general and complex form than sequential recommendation which recommends the next item according to the historical records. However, the majority of recent literature fails to simultaneously consider the two aspects: 1) complex order-invariant relation of multiple elements within a set; 2) intangible order-variant user appetite in users' historical interactions. To overcome the barrier, we propose a novel Conv-based Temporal Sets Prediction model (dubbed ConvTSP). We contrive a Conv-based Relationship Learning module to capturing the cross-relation between items in each basket. Besides, we integrate sequential-level dynamic user appetite and individual-level static user appetite into a fused user appetite representation. Extensive experiments and ablation studies are done to show how effective ConvTSP is.