Multi-Intent Driven Contrastive Sequential Recommendation

Abstract

Sequential Recommendation (SR) models with auxiliary tasks of contrastive learning have achieved remarkable progress in recent years, which can effectively mine the self-supervised signals to mitigate the data sparsity problem. However, current contrastive SR models overlook the intricate correlations among different users, leading to the false negative pair problem and adversely affecting recommendation performance. Therefore, in this paper, we propose a multi-intent driven contrastive SR model MICRec. MICRec learns global intent prototypes from the users by a moving-average updating strategy. Then, MICRec introduces two multi-intent guided contrastive losses, including a sequence-level contrastive loss and an intent-level contrastive loss, which both contribute to mining the self-supervised information and building accurate user embeddings. The former optimizes the negative sample set by eliminating the false negative sequence pairs with overlapping intents, and the latter further stabilizes the latent structure by aligning the intents excavated from the original and augmented sequences. Thus, with the multi-intent guided contrastive learning strategy, our model can better understand the correlations between users, leading to a more effective and accurate representation structure in the latent space. MICRec not only achieves superior performance, but also improves the robustness to the interaction noise. The experimental results on three public benchmark datasets show that MICRec outperforms existing SR models in terms of Recall and NDCG.