Knowledge-Aware Intent-Guided Contrastive Learning for Next-basket Recommendation

Next-Basket Recommendation (NBR) aims to predict a series of items in the next basket based on users' current basket sequence. However, the existing works merely consider the explicit auxiliary signals, and intent may contribute to the refinement of basket representations but bring some uncertain bias.To deal with the problems mentioned above, this paper proposes a knowledge-aware intent-guided contrastive learning method called KICL for NBR. Specifically, we construct a collaborative bipartite graph to learn basket representations and item representations, while at the same time, a knowledge graph is constructed based on items and their attributes to capture implicit auxiliary signals. Furthermore, the item attributes within a basket are weighted and summed up to extract the corresponding intent. To reduce the uncertainty bias brought from item diversity, a contrastive regularizer is designed for better basket representation refinement. Extensive experiments on two real-world datasets demonstrate the effectiveness of KICL, where the maximum improvement can reach 15.91% in terms of F1@10 on Dunnhumby.