GraphFusionSBR: Denoising Multi-Channel Graphs for Session-Based Recommendation
Jia-Xin He, Hung-Hsuan Chen · 2026
Read GraphFusionSBR: Denoising Multi-Channel Graphs for Session-Based Recommendation by Jia-Xin He, Hung-Hsuan Chen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-08497.txt.
Introduction
Jia-Xin He and Hung-Hsuan Chen present GraphFusionSBR for session-based recommendation. Their 2026 arXiv paper notes that session-based systems must infer user intent from short click sessions. Existing models suffer from item interaction dominance and noisy sessions. The authors propose a multi-channel model that denoises across three graph views. Channels include a knowledge graph, a session hypergraph, and a session line graph. Fusing several views aims to capture intent that any single graph would miss, improving suggestions for anonymous shoppers.
Catalog record
| CREATOR | Jia-Xin He, Hung-Hsuan Chen |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 68 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-08497
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
GraphFusionSBR: Denoising Multi-Channel Graphs for Session-Based Recommendation Jia-Xin He, Hung-Hsuan Chen 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. 1 1 1 https://github.com/hohehohe0509/DSR-HK K eywords Session-Based Recommendation \cdot Graph Neural Networks \cdot Knowledge Graph \cdot Hypergraph \cdot Self-Supervised Learning 1 Introduction Recommendation systems have become an integral part of our digital lives, influencing the way we consume information, entertainment, and products. These systems play a crucial role in various domains, including e-commerce, social media, and content streaming platforms, by tailoring user experiences based on their preferences. Traditional recommendation systems primarily rely on user profiles, past interactions, or historical browsing data to suggest items. Although effective in specific contexts, these systems face significant challenges, especially in dynamic environments where user preferences are volatile or complex to capture. One of the main challenges in traditional recommendation systems is privacy. With increasing concerns over data protection, long-term user profiles, which are typically required by these systems, can no longer be freely utilized due to stringent privacy regulations, such as GDPR. This limits the potential of personalized recommendations and imposes restrictions on data collection practices. Additionally, the cold start problem, where a system struggles to recommend items to new users or for new items due to a lack of sufficient historical data, remains a substantial hurdle. Furthermore, user preferences can change over time, with different needs and interests emerging across different browsing sessions. Traditional models, which are based on static user profiles, often fail to capture these session-based short-term preferences. In these contexts, session-based recommendation (SBR) ( 7 ) has emerged as a practical approach. Unlike traditional systems, SBR models capture users’ immediate preferences during a single browsing session, without requiring any long-term user information. This makes SBR inherently more privacy-friendly, as it circumvents the need for extensive user data collection. Furthermore, SBR is highly suitable for real-time recommendation tasks, where quick adaptability to evolving user interests is crucial. By focusing on short-term preferences, SBR not only mitigates the cold start problem but also likely enhances the user experience by providing timely and relevant suggestions. Existing SBR systems can be broadly categorized into three main types: traditional models, sequence-based models, and graph neural network (GNN) models ( 6 ) . Conventional models, such as those based on Markov chains, introduce temporal dependencies by capturing transitions between adjacent items in a session. However, these models often fail to account for more complex, higher-order relationships between items. Sequence-based models ( 7 ; 9 ; 12 ) use recurrent neural networks (RNNs) and their variants, such as the Gated Recurrent Unit (GRU) ( 4 ) , to model sequences of items in temporal order. These models have proven effective in learning sequential patterns but may struggle with long-range or complex, high-order dependencies due to the inherent limitations of RNNs. More recently, graph neural networks (GNNs) have been increasingly applied to SBR because they allow sessions to be modeled as graphs, where nodes represent items and edges represent interactions between them. GNN-based models offer a robust framework for capturing complex item relationships, as they can naturally represent both direct and indirect dependencies. These methods can be further divided into two categories: ordinary graph-based models and hypergraph-based models. Ordinary graphs capture pairwise item interactions within a session, while hypergraphs connect all items in a session, enabling the capture of higher-order item relationships that are often crucial in modeling user intent more comprehensively. Despite these advances, several challenges remain in the development of effective SBR systems. First, existing models often fail to leverage external information related to the items, such as knowledge graphs, which could provide valuable context by modeling semantic relationships between items. Knowledge graphs can enrich the recommendation process by incorporating domain knowledge, but they are often underutilized due to the dominance of direct item interactions. Furthermore, even when knowledge graphs are integrated, their impact can be overshadowed by noisy or irrelevant data, resulting in poor encoding of valuable external information into item representations. Lastly, noise within the session data itself – arising from factors such as user misclicks or browsing anomalies – can further degrade the performance of the recommendation system by introducing distractions from the user’s true intent. To address these issues, we propose a novel SBR model that leverages the strengths of multiple graph-based representations while mitigating noise. Our model, called GraphFusionSBR, introduces a multi-channel approach that integrates three types of graphs: a knowledge graph, a session hypergraph, and a session line graph. Each channel is designed to capture different aspects of user interaction. The knowledge graph channel incorporates external knowledge to enrich item representations, while a view generator dynamically removes irrelevant or redundant edges, reducing the risk of noise. In the session hypergraph channel, all session items are interconnected, enabling the model to capture complex inter-dependencies among items. To address noise within sessions, the session line graph channel assigns attention scores based on item similarity, allowing the model to focus on relevant items and generate denoised session representations. These representations are further enhanced through cross-channel contrastive learning ( 2 ) , which maximizes mutual information between the different channels, leading to more robust and accurate recommendations. Unlike existing multi-view approaches that simply concatenate representations from different graphs, GraphFusionSBR introduces a selective fusion mechanism centered on denoising. We argue that more information does not always lead to better recommendations; rather, filtering noise from external knowledge and internal session interactions is key. The main contributions of this work are summarized as follows: • We propose a novel unified multi-channel denoising framework that integrates knowledge graphs (from knowledge base), hypergraphs (from session information), and line graphs (from session information). Crucially, we incorporate a view-generator mechanism in the knowledge graph channel and an importance extraction module in the line graph channel to actively filter out irrelevant semantic edges and noisy session interactions before fusion. • We bridge the gap between semantic-based reasoning (via knowledge graph) and structure-based reasoning (via hypergraph and line graph). By maximizing mutual information between these views through cross-channel contrastive learning, we enhance the robustness of item representations against data sparsity. • Extensive experiments on three benchmark datasets (Tmall, RetailRocket, and KKBox) demonstrate that GraphFusionSBR consistently outperforms state-of-the-art baselines. Furthermore, our theoretical analysis confirms that this performance gain comes with linear computational complexity, making it scalable for real-world applications. The remainder of the paper is organized as follows. In Section 2 , we review previous work on session-based recommendation, knowledge graph, hypergraph, and hypergraph-based recommendation, and self-supervised learning on recommendation systems. Section 3 introduces our proposed method.