Diversifying Collaborative Filtering via Graph Spreading Network and Selective Sampling.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Graph Neural Networks (GNNs) excel at processing graph data and achieving high accuracy in collaborative filtering (CF) recommendations.
• Existing GNN-based recommendation models face an accuracy-diversity dilemma, where increasing recommendation diversity significantly degrades accuracy.
• Current GNN models lack flexibility in adjusting the trade-off between accuracy and diversity for varied user demands.

Purpose of the Study:

• To address the accuracy-diversity dilemma in GNN-based recommendations.
• To develop a flexible GNN model that can adapt to different accuracy-diversity ratio requirements.
• To improve the diversity of recommendations while maintaining or enhancing accuracy.

Main Methods:

• Proposed Graph Spreading Network (GSN), a novel GNN model utilizing neighborhood aggregation for CF.
• GSN employs both diversity-oriented and accuracy-oriented aggregations during embedding propagation.
• Introduced a new sampling strategy for selecting accurate and diverse negative samples to aid model training.

Main Results:

• GSN effectively resolves the accuracy-diversity dilemma, achieving improved recommendation diversity with maintained accuracy.
• The model demonstrated significant improvements over state-of-the-art methods, with average increases in R@20, N@20, G@20, and E@20 across three real-world datasets.
• A hyper-parameter in GSN allows for tunable control over the accuracy-diversity ratio, catering to specific application needs.

Conclusions:

• GSN offers a robust solution for enhancing recommendation diversity in collaborative filtering.
• The model's flexibility in balancing accuracy and diversity makes it suitable for various recommendation scenarios.
• GSN represents a significant advancement in developing more balanced and adaptable GNN-based recommendation systems.