Multi-channel Contrastive Learning for Sequential Recommendation

Abstract

The purpose of Sequential Recommendation (SR) is to recommend the next commodities that a user wants to buy based on their historical interaction sequence. The current approach for SR focuses only on mining the user’s interest preferences, while they all fail to consider the influence of item prices on users’ purchase decisions and suffer from the data sparsity problem. In this paper, a Multi-channel Contrastive Learning method for SR (MCLSR) is proposed, which can effectively extract users’ interest preferences and price preferences and alleviate the sparsity issues. Specifically, first, a heterogeneous knowledge graph is constructed from all interaction sequence and the item attribute (i.e., item price and category) by us. Then, we leverage a heterogeneous graph neural network mechanism to learn user, item, and price node embeddings. Next, users’ price preferences and interest preferences are extracted by an attention network. Finally, a multi-channel contrastive learning mechanism is employed to build price and interest preferences’ relations and generate high-quality recommendation results. Experiments on both real datasets show that MCLSR obtains more sophisticated performance than the existing baseline.