Dual-level graph contrastive collaborative filtering.

Area of Science:

• Artificial Intelligence
• Computer Science

Background:

• Graph-based collaborative filtering is effective for recommendation systems but struggles with data sparsity.
• Existing contrastive learning methods in graph collaborative filtering use single-view designs, limiting performance.

Purpose of the Study:

• To propose a novel dual-level graph contrastive collaborative filtering method (DL-GCL).
• To enhance recommendation system robustness and mitigate data sparsity by integrating both graph and node-level views.

Main Methods:

• DL-GCL employs matrix decomposition for node-level contrastive view construction.
• Graph-level contrastive views are generated using the Fast Gradient Sign Method (FGSM) to mitigate noise.
• A multi-task learning strategy optimizes local-global views and Bayesian Personalized Ranking (BPR) loss.

Main Results:

• Experiments on four datasets show significant improvements in recommendation performance.
• NDCG and Recall metrics improved by up to 24.5% compared to state-of-the-art graph contrastive models.

Conclusions:

• DL-GCL effectively improves recommendation system robustness.
• The dual-level contrastive approach successfully mitigates data sparsity issues.