Learn۲Rank: Recommender System through Adaptive Contrastive Learning

The ranking is an essential part of recommender systems. While traditional collaborative filtering effectively identifies user-item similarities, it struggles with cold starts and data sparsity. To address these issues, we propose a novel recommender system that employs a heterogeneous Graph Attention Network combined with adaptive contrastive learning to improve representation learning and ranking accuracy. We model user-item interactions as a heterogeneous graph with bidirectional edges for comprehensive information propagation. By integrating contrastive learning with Bayesian Personalized Ranking (BPR) loss, our model effectively differentiates relevant from irrelevant items, enhancing recommendation precision. Also, we use an ensemble approach to aggregate predictions from multiple independently trained models, reducing variance and mitigating overfitting. This study demonstrates the effectiveness of combining heterogeneous graph neural networks with self-supervised learning and ensemble techniques in advancing the state-of-the-art in recommender systems.

کلیدواژه ها:

Recommender System ، Ranking ، Adaptive Contrastive Learning