Efficient Graph Collaborative Filtering via Contrastive Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Graph neural networks (GNNs) are effective for collaborative filtering (CF) but suffer from inefficient training and biased information propagation.
• Existing methods struggle with sparse user-item interactions, limiting the effectiveness of traditional Bayesian personalized ranking (BPR) loss.

Purpose of the Study:

• To propose an Efficient Graph Collaborative Filtering (EGCF) method that addresses the limitations of existing GNN-based CF approaches.
• To enhance recommendation system performance through improved training efficiency and representation learning.

Main Methods:

• EGCF utilizes a single-layer graph convolution for modeling user-item interactions from first-order neighbors.
• Contrastive learning is incorporated to derive self-supervisions, enhancing user and item representation learning.
• The model is jointly trained with supervised learning signals.

Main Results:

• EGCF achieves state-of-the-art performance on Yelp2018 and Amazon-book datasets, particularly in Recall and NDCG metrics.
• The proposed method demonstrates significant advantages in training efficiency compared to baseline models.
• Experimental results confirm EGCF's effectiveness in accurately ranking target items.

Conclusions:

• EGCF offers a more efficient and effective approach to graph-based collaborative filtering.
• The integration of contrastive learning improves representation learning and recommendation accuracy.
• EGCF presents a practical solution for real-world recommendation applications due to its efficiency and performance.