Core Concepts
Intent-aware Recommendation via Disentangled Graph Contrastive Learning (IDCL) enhances recommender systems by disentangling user intents and inferring behavior distributions.
Abstract
Abstract:
Graph neural networks (GNN) are popular for recommender systems.
Understanding user intents is crucial for recommender systems.
IDCL simultaneously learns interpretable intents and behavior distributions.
Introduction:
Recommender systems alleviate information overload.
GNN-based systems explore multi-hop relationships for better representation.
Methodology:
Behavior Disentangling module disentangles user intents.
Intent-wise Contrastive Learning enhances disentangling and infers behavior distributions.
Coding Rate Reduction Regularization promotes independence of behaviors across different intents.
Experiment:
IDCL outperforms SOTA baselines in recommendation performance.
Independence analysis shows that different intents are independent.
Explainability analysis demonstrates the interpretability of learned representations.
Stats
Graph neural networks (GNN) based recommender systems are mainstream.
IDCL substantially improves recommendation performance.
IDCL disentangles user intents and infers behavior distributions.