A Graph-Neural-Network-Based Social Network Recommendation Algorithm Using High-Order Neighbor Information.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Social-network-based recommendation systems utilize social connections to improve recommendations.
• Traditional methods often overlook high-order collaborative signals, limiting performance.
• Data sparsity remains a challenge in recommendation systems.

Purpose of the Study:

• To propose a novel graph neural network (GNN)-based social recommendation model.
• To effectively capture and leverage high-order collaborative signals.
• To enhance recommendation performance by integrating user-item interactions and social network data.

Main Methods:

• Developed a GNN framework to learn latent representations of users and items.
• Stacked embedding propagation layers to aggregate multi-hop neighborhood information.
• Utilized both user-item interaction and social network graphs.
• Employed a lightweight GNN by retaining neighborhood aggregation and omitting feature transformation/nonlinear activation.

Main Results:

• The proposed GNN-based model successfully captures high-order collaborative signals.
• Explicitly injected collaborative signals from both interaction and social graphs into entity representations.
• Demonstrated superior performance compared to state-of-the-art recommendation algorithms on two real-world datasets.
• The lightweight GNN framework eased training and alleviated overfitting.

Conclusions:

• The GNN-based social recommendation model offers a significant improvement over existing methods.
• Leveraging high-order collaborative signals through GNNs is crucial for effective social recommendation.
• The proposed approach provides a robust and efficient solution for data sparsity and recommendation accuracy.