Session-based recommendation aims to recommend the next item based on short-term interactions. Traditional session-based recommendation methods assume that all interacted items are closely related to the user's interests. However, noise (e.g., accidental clicks) is inevitably present in the interaction data. Existing methods for session-based recommendation only focus on graph denoising for enhanced item embeddings, resulting in sub-optimal session representation learning. To address these issues, we propose RAIN: Reconstructed-Aware In-context eNhancement with Graph Denoising for session-based recommendation. RAIN performs denoising on both the graph and session in a step-by-step manner. Guided by self-supervised signals, we aim to enhance the clarity of edges by employing masking and reconstruction alongside training an edge indicator to effectively eliminate noisy edges. By leveraging the trained indicator and incorporating a self-attentive mechanism, we additionally incorporate reconstructed-aware in-context enhancement within the session. Comparative evaluations with current state-of-the-art methods demonstrate that RAIN achieves significant improvements, with gains up to 7.05% in Hit@20 and 1.53% in MRR@20 on four benchmark datasets. The experimental results and analysis provide evidence for the rationality and superiority of our proposed model. The source code is available at https://github.com/zengxy20/RAIN.
Keywords: Graph neural networks; Self-supervised learning; Session denoising; Session-based recommendation.
Copyright © 2024. Published by Elsevier Ltd.