Debiased Sequential Recommendation via Multi-intent Disentanglement and Conformity-Aware Contrastive Learning

The goal of sequential recommendation is to achieve personalized recommendations by analyzing the user's historical interaction sequence and capturing the dynamically changing interest preferences. However, this process is often influenced by popularity bias, where popular items are recommended more frequently. Conformity, the user's tendency toward popular items, is an important factor contributing to this issue. The existing recommendation models have not fully considered the differences in conformity between users’ multiple intentions, failing to make recommendations based on the consistency preferences of different intents, resulting in reduced performance. Moreover, a user's multiple intentions are often entangled, making it difficult to distinguish the differences in conformity between different intentions. Motivated by this issue, we design a Multi-Intent Disentanglement and Conformity-Aware Contrastive Learning method (MIDCRec). Specifically, MIDCRec first disentangles the user's historical interactions into multiple intents and extracts the user's current primary intent. To measure the tendency of the primary intent towards popular items, a conformity-aware module is further proposed to extract self-supervised signals of intent conformity. Finally, based on the conformity of the dominant intent, an adaptive contrastive learning loss weighting method is proposed to mitigate the interference of popularity bias on recommendations. Experimental results on four real-world datasets confirm the effectiveness of the proposed method.