RAIN: Reconstructed-aware in-context enhancement with graph denoising for session-based recommendation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Session-based recommendation systems predict user interests from short-term interactions.
• Traditional methods struggle with noisy data (e.g., accidental clicks), leading to suboptimal session representations.
• Existing approaches primarily focus on graph denoising for item embeddings, neglecting session-level noise.

Purpose of the Study:

• To propose RAIN (Reconstructed-Aware In-context eNhancement with Graph Denoising), a novel model for session-based recommendation.
• To address limitations in session representation learning caused by noisy interaction data.
• To enhance recommendation accuracy by denoising both the interaction graph and the user session.

Main Methods:

• RAIN employs a step-by-step denoising process for both the graph and session data.
• Self-supervised signals guide edge clarity enhancement through masking and reconstruction.
• An edge indicator is trained to identify and eliminate noisy edges, improving graph structure.
• Reconstructed-aware in-context enhancement is integrated using a self-attentive mechanism and the trained edge indicator.

Main Results:

• RAIN achieved significant improvements over state-of-the-art methods on four benchmark datasets.
• Performance gains reached up to 7.05% in Hit@20 and 1.53% in MRR@20.
• Experimental analysis validated the model's rationality and superiority in session-based recommendation.

Conclusions:

• The proposed RAIN model effectively handles noisy data in session-based recommendation.
• RAIN enhances both item embeddings and session representations for improved accuracy.
• The method offers a superior approach to session-based recommendation compared to existing techniques.