Graph Collaborative Filtering Based on Dual-Message Propagation Mechanism.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Collaborative Filtering (CF) is a key approach in recommender systems, leveraging user-item preferences and user/item similarities.
• Graph Neural Networks (GNNs) have enhanced CF models, but often only implicitly utilize item similarities, limiting performance.
• Existing GNN-based CF models struggle to explicitly model both user-item preferences and item similarities simultaneously.

Purpose of the Study:

• To address the limitation of implicit similarity modeling in GNN-based CF.
• To propose a novel mechanism that explicitly captures both user-item preferences and item similarities.
• To develop a new GNN-based CF model that better aligns with the core philosophy of collaborative filtering.

Main Methods:

• Introduction of a Dual-Message Propagation (DPM) mechanism.
• Development of a Dual-Message Graph Collaborative Filtering (DGCF) model utilizing DPM.
• Each user/item embedding is processed by two GNNs: one for preferences, one for similarities.

Main Results:

• The DGCF model significantly outperforms state-of-the-art CF models on three real-world datasets.
• The proposed DPM mechanism effectively models both preferences and similarities.
• Experimental results validate the superiority of explicitly modeling both relationship types.

Conclusions:

• The DGCF model offers a substantial improvement in recommendation performance.
• Explicitly modeling user-item preferences and item similarities is crucial for advanced CF.
• The DGCF model represents a more complete realization of the collaborative filtering philosophy.