HYBRID DISENTANGLED GRAPH CONTRASTIVE LEARNING FOR INTENT-AWARE RECOMMENDATION

Abstract: Recommender systems play a vital role in modern digital platforms by delivering personalized content to users. However, capturing the dynamic and multifaceted nature of user intent remains a significant challenge. Traditional models rely on static user-item interactions and fail to disentangle the multiple latent factors that drive user behavior. This paper proposes a Hybrid Disentangled Graph Contrastive Learning (HDGCL) framework for intent-aware recommendation. The model constructs a user-item interaction graph and applies disentangled representation learning to separate latent factors such as long-term preferences, short-term behavioral patterns, and situational context. A hybrid contrastive learning mechanism is employed to enhance robustness and discriminability of learned embeddings. Contextual signals including time, location, and mood are incorporated to enable dynamic adaptation of user intent. Experimental results demonstrate that HDGCL consistently outperforms state-of-the-art recommendation baselines in Precision, Recall, and NDCG while improving diversity and interpretability.

Keywords: Hybrid Disentangled Graph Contrastive Learning, Intent-Aware Recommendation, Graph Neural Networks, Disentangled Representation Learning, Contrastive Learning, User Behavior Modeling, Context-Aware Recommendation