A Hyperbolic Graph Neural Network Model with Contrastive Learning for Rating-Review Recommendation

Area of Science:

• Recommender Systems
• Graph Neural Networks
• Machine Learning

Background:

• Data sparsity is a major challenge in recommender systems, limiting accuracy.
• Conventional graph neural networks (GNNs) struggle with non-Euclidean data structures, hindering performance.
• Existing methods often fail to capture heterogeneous interaction patterns effectively.

Purpose of the Study:

• To propose a hyperbolic graph neural network model for rating-review recommendation.
• To enhance recommendation accuracy by integrating multimodal information, specifically reviews and user-item interactions.
• To address the limitations of Euclidean embeddings in modeling complex real-world data.

Main Methods:

• Implemented a dual-graph construction: a review-aware graph and a user-item interaction graph.
• Utilized a hyperbolic graph neural network architecture for joint high-order feature learning.
• Incorporated contrastive learning in hyperbolic space to leverage semantic and interaction data.

Main Results:

• The proposed model significantly improves recommendation accuracy on real-world datasets.
• Demonstrated superior performance compared to conventional methods in handling data sparsity.
• Effectively avoided embedding distortion issues common in high-order feature learning.

Conclusions:

• The hyperbolic graph neural network with contrastive learning offers a powerful approach for rating-review recommendation.
• Integrating review semantics and user-item interactions in hyperbolic space enhances representation capacity.
• The model provides a robust solution for improving recommender system performance and accuracy.