Intent-enhanced sequential recommendation via multi-view contrastive learning over stabilized essential intents

Abstract

Sequential recommendation provides an effective way to model users' historical behaviors for predicting future interactions. However, the existing methods have not fully explored how to conduct stable intent data augmentation based on user sub-intents, and simultaneously construct multi-view self-supervised signals for intent contrastive learning to assist in the next step of prediction. To address these limitations, we propose IeSRec, an intent-enhanced sequential recommendation model. It explicitly incorporates the original behavioral intentions of users by introducing three novel data augmentation strategies. These strategies construct diverse behavioral subsequences from multiple perspectives, thereby strengthening the representation of implicit intents. Based on operations such as segmentation, encoding, and clustering, we design a multi-view self-supervised framework to capture distinctions among different user intentions. Additionally, we formulate three contrastive learning objectives to bring users with similar intentions closer in the embedding space while pushing those with dissimilar intentions apart. The intent contrastive learning module is jointly optimized with the primary next-item prediction task in an end-to-end manner. Extensive experiments on six public datasets demonstrate that IeSRec consistently outperforms state-of-the-art baselines. Ablation studies further validate the effectiveness of the proposed data augmentation methods, and their portability to other sequential recommendation architectures.