GraphFusionSBR: Denoising Multi-Channel Graphs for Session-Based Recommendation

TL;DR

GraphFusionSBR addresses noise and data sparsity in session-based recommendations by fusing three denoised graph views: a knowledge-graph channel with a view-generator, a session hypergraph channel, and a session line graph channel. The model jointly optimizes $L = L_{rec} + \lambda_1 L_{ssl} + \lambda_2 L_{KG}$, leveraging cross-channel contrastive learning to enrich embeddings while mitigating noise. Key innovations include a KG denoising module, an Importance Extraction Module for the line graph, and a cross-view SSL objective that ties semantic and structural signals across channels. Across Tmall, RetailRocket, and KKBox, GraphFusionSBR achieves state-of-the-art performance with linear scalability, demonstrating strong practical potential for diverse domains.

Abstract

Session-based recommendation systems must capture implicit user intents from sessions. However, existing models suffer from issues such as item interaction dominance and noisy sessions. We propose a multi-channel recommendation model, including a knowledge graph channel, a session hypergraph channel, and a session line graph channel, to capture information from multiple sources. Our model adaptively removes redundant edges in the knowledge graph channel to reduce noise. Knowledge graph representations cooperate with hypergraph representations for prediction to alleviate item dominance. We also generate in-session attention for denoising. Finally, we maximize mutual information between the hypergraph and line graph channels as an auxiliary task. Experiments demonstrate that our method enhances the accuracy of various recommendations, including e-commerce and multimedia recommendations. We release the code on GitHub for reproducibility.\footnote{https://github.com/hohehohe0509/DSR-HK}

Figures (4)

• Figure 1: The overall architecture of GraphFusionSBR. The raw session data and knowledge base are processed through three parallel channels. The knowledge graph channel (top) integrates a knowledge base and applies a denoising view generator. The hypergraph channel (middle) captures high-order correlations. The line graph channel (bottom) extracts cross-session information via an Importance Extraction Module (IEM). The final representations are fused for next-item prediction, optimized jointly by recommendation loss, cross-channel contrastive loss, and knowledge graph loss.
• Figure 2: Effect of different layers on performance across datasets
• Figure 3: Effect of different $\lambda_1$s on performance across datasets
• Figure 4: Effect of different $K$s on performance across datasets