Enhancing Sequential Recommendation via Aligning Interest Distributions

Contrastive learning improves the performance of sequential recommendation models by mining self-supervised information and mitigating the impact of data sparsity and noise interference. Existing contrastive sequential recommendation models pull the embeddings of positive sequence pairs close, and train sequence encoders to be invariant to data augmentations, e.g., reordering, which could destroy information beneficial for the recommendation task, e.g., the order of interactions. To alleviate the problem, we propose a contrastive sequential recommendation model IDARec, which adds projection heads between the sequence encoder and contrastive loss and builds the recommendation loss and contrastive loss in different hidden spaces. Specifically, IDARec introduces an interest distribution-based contrastive loss, which transforms sequence embeddings into multi-grained interest distributions and aligns the interest distributions of positive sequence pairs. Moreover, a clustering-classification approach is adopted to learn interest distributions, which learns interest prototypes by K-means first and then interest distributions by classification. We conduct extensive experiments on four public datasets, and the experimental results show that our model outperforms the state-of-the-art sequential recommendation models.