A graph neural network recommendation algorithm based on multi-scale attention and contrastive learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Recommendation systems struggle with sparse user-item interaction data, hindering representation learning and overall performance.
• Existing methods often fail to effectively model higher-order user-item relationships or mitigate biases in graph structures.

Purpose of the Study:

• To propose a novel graph neural network-based recommendation algorithm (GR-MC) designed to address challenges posed by sparse data.
• To enhance the quality of user and item representations for improved recommendation accuracy and robustness.

Main Methods:

• Implemented a graph structure augmentation strategy using user-focused edge dropout to reduce degree bias.
• Introduced a multi-scale attention mechanism for improved embedding propagation and modeling of higher-order relationships.
• Incorporated contrastive learning as a self-supervised task to enhance embedding discriminative ability and model robustness.

Main Results:

• GR-MC demonstrated superior performance compared to existing methods across multiple public datasets.
• Achieved a significant 24.69% improvement in Recall@20 on the highly sparse Amazon-book dataset.
• Validated the model's effectiveness and robustness, particularly in environments with limited interaction data.

Conclusions:

• The proposed GR-MC algorithm effectively overcomes the limitations of sparse user-item interaction data.
• Multi-scale attention and contrastive learning significantly enhance representation learning and recommendation performance.
• GR-MC offers a robust solution for recommendation systems operating in data-scarce scenarios.