Intent-Interest Disentanglement and Item-Aware Intent Contrastive Learning for Sequential Recommendation

TL;DR

This paper introduces IDCLRec, a novel sequential recommendation model that disentangles user behaviors into intents and interests, and employs item-aware contrastive learning to improve personalized recommendations by capturing dynamic motivations and stable tastes.

Contribution

The paper proposes a new method for intent-interest disentanglement and item-aware contrastive learning that does not require predefined intent categories, enhancing understanding of user behaviors in recommender systems.

Findings

IDCLRec outperforms existing models on real-world datasets.

Disentangling intents and interests improves recommendation accuracy.

Item-aware contrastive learning enhances user behavior modeling.

Abstract

Recommender systems aim to provide personalized item recommendations by capturing user behaviors derived from their interaction history. Considering that user interactions naturally occur sequentially based on users' intents in mind, user behaviors can be interpreted as user intents. Therefore, intent-based sequential recommendations are actively studied recently to model user intents from historical interactions for a more precise user understanding beyond traditional studies that often overlook the underlying semantics behind user interactions. However, existing studies face three challenges: 1) the limited understanding of user behaviors by focusing solely on intents, 2) the lack of robustness in categorizing intents due to arbitrary fixed numbers of intent categories, and 3) the neglect of interacted items in modeling of user intents. To address these challenges, we propose Intent-Interest Disentanglement and Item-Aware Intent Contrastive Learning for Sequential Recommendation (IDCLRec). IDCLRec disentangles user behaviors into intents which are dynamic motivations and interests which are stable tastes of users for a comprehensive understanding of user behaviors. A causal cross-attention mechanism is used to identify consistent interests across interactions, while residual behaviors are modeled as intents by modeling their temporal dynamics through a similarity adjustment loss. In addition, without predefining the number of intent categories, an importance-weighted attention mechanism captures user-specific categorical intent considering the importance of intent for each interaction. Furthermore, we introduce item-aware contrastive learning which aligns intents that occurred the same interaction and aligns intent with item combinations occurred by the corresponding intent. Extensive experiments conducted on real-world datasets demonstrate the effectiveness of IDCLRec.