Simplified self-supervised learning for hybrid propagation graph-based recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Graph Convolutional Networks (GCNs) show promise in recommendation systems.
• Existing GCN methods face challenges in leveraging multi-order connectivity, augmenting sparse data, and mitigating noise.

Purpose of the Study:

• To develop a novel GCN-based recommendation method addressing limitations of current approaches.
• To improve the effectiveness and robustness of GCNs in recommendation tasks.

Main Methods:

• Introduced a hybrid propagation GCN framework (S3HGN) with nonlinear propagation.
• Incorporated a simplified self-supervised learning paradigm for contrastive view generation.
• Employed residual prediction through weighted summation and dropout for noise mitigation.

Main Results:

• S3HGN effectively leverages multi-order graph connectivity for improved node embeddings.
• The self-supervised strategy enhances model robustness against noisy data.
• Experiments demonstrate superior performance compared to eight representative graph-based collaborative filtering models.

Conclusions:

• S3HGN offers an effective and robust solution for GCN-based recommendation.
• The proposed hybrid propagation and self-supervised learning approach addresses key challenges in the field.