Noise-augmented contrastive learning with attention for knowledge-aware collaborative recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Knowledge graphs (KGs) are crucial for recommender systems, with Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) dominating Collaborative Knowledge Graph (CKG) models.
• Large-scale recommender systems face challenges with long-tail distributions and data sparsity, leading to uneven entity embedding distributions.
• Contrastive Learning (CL) helps mitigate data sparsity by learning general representations, but traditional graph augmentation techniques are suboptimal for CL-based recommendations.

Purpose of the Study:

• To propose a novel method, Noise Augmentations Knowledge Graph Attention Contrastive Learning (NA-KGACL), to improve recommender systems.
• To address the limitations of existing graph augmentation techniques in CL-based recommendations.
• To enhance the handling of long-tail distributions and data sparsity in large-scale graph-based recommender systems.

Main Methods:

• Developed a multi-level contrastive framework integrating CL with Knowledge-GAT.
• Refined node representations using projection heads and shuffled batch normalization.
• Introduced a noise-augmented algorithm as a replacement for ineffective graph augmentation methods to generate contrastive learning views.

Main Results:

• The proposed NA-KGACL method demonstrated improved learned representations on three large-scale, real-world datasets.
• Experimental results showed increased recommendation accuracy compared to existing methods.
• The study indicated more efficient training processes with the NA-KGACL approach.

Conclusions:

• NA-KGACL effectively addresses data sparsity and long-tail distribution issues in graph-based recommender systems.
• The noise augmentation strategy provides a powerful alternative for generating contrastive views.
• The method offers significant improvements in both recommendation performance and training efficiency.