Contrastive Learning Method for Sequential Recommendation based on Multi-Intention Disentanglement

TL;DR

This work tackles the challenge of modeling multiple dynamic user intentions in sequential recommendation. It proposes MIDCL, a Transformer-based framework that disentangles user intents with a Variational Auto-Encoder and enhances representations via two contrastive learning paths: intention-based Triplet loss and sequence-based InfoNCE loss. Empirical results on four real-world datasets show MIDCL achieves state-of-the-art performance while offering interpretability of the latent intents, as evidenced by intention visualizations. The approach advances robust, explainable sequential recommendations by explicitly handling irrelevant intentions and leveraging self-supervised signals.

Abstract

Sequential recommendation is one of the important branches of recommender system, aiming to achieve personalized recommended items for the future through the analysis and prediction of users' ordered historical interactive behaviors. However, along with the growth of the user volume and the increasingly rich behavioral information, how to understand and disentangle the user's interactive multi-intention effectively also poses challenges to behavior prediction and sequential recommendation. In light of these challenges, we propose a Contrastive Learning sequential recommendation method based on Multi-Intention Disentanglement (MIDCL). In our work, intentions are recognized as dynamic and diverse, and user behaviors are often driven by current multi-intentions, which means that the model needs to not only mine the most relevant implicit intention for each user, but also impair the influence from irrelevant intentions. Therefore, we choose Variational Auto-Encoder (VAE) to realize the disentanglement of users' multi-intentions. We propose two types of contrastive learning paradigms for finding the most relevant user's interactive intention, and maximizing the mutual information of positive sample pairs, respectively. Experimental results show that MIDCL not only has significant superiority over most existing baseline methods, but also brings a more interpretable case to the research about intention-based prediction and recommendation.

Figures (6)

• Figure 1: An example about historical behavioral sequences of different users, each item represents a kind of interactive behavior.
• Figure 2: Overview of MDICL. (a) shows the basic structure of Transformer encoder. (b) describes the framework and process of MIDCL, which includes the intention disentanglement by VAE and contrastive learning. The detailed illustration of VAE is in Sect.\ref{['sec4.3']}, and five kinds of augmentation operators is recorded in Sect.\ref{['sec4.4']}. Multi-task training integrates adaptive results from three types of loss functions for the final prediction.
• Figure 3: Performance comparison w.r.t. Hyper-Parameter $\alpha$.
• Figure 4: Performance comparison w.r.t. Hyper-Parameter $\beta$.
• Figure 5: Performance comparison w.r.t. the number of intentions ${k}$ where $\alpha$ is 0.1 and $\beta$ is 0.3.