Learning unified denoised representations for sequential recommendation

Abstract

Sequential Recommendation (SR) has been widely used in many internet applications, such as e-commerce, social platforms and hot news. User behavior sequential data in SR typically contains complex patterns of short-term dependencies, long-term dependencies and noise, which may lead SR models to misinterpret user intentions and overfit noisy patterns, but existing methods cannot solve them simultaneously. To address this problem, we propose a model to learn Unified Denoised Representations (UniDR) for the SR task, which consists of three modules. The first module employs Graph Neural Networks (GNNs) with adaptive learning mechanisms to capture short-term dependencies by dynamically weighting item transitions. The second module utilizes self-attention mechanisms to effectively model long-term dependencies across the entire item sequence. The third module focuses on extracting long-term sequential patterns from contextual information through feed-forward networks. Each module independently generates denoised representations, either through weak edge removal in GNNs or through frequency domain transformations. UniDR integrates these denoised representations by jointly optimizing a BPR loss, an alignment loss and a uniformity loss. Extensive experiments on five public benchmark datasets demonstrate UniDR’s superiority in recommendation performance and robustness to interaction noise. Compared to the strongest state-of-the-art baseline, UniDR achieves significant improvements, with average increases of 10.63% in Hit Rate (HR), 21.47% in Normalized Discounted Cumulative Gain (NDCG) and 23.71% in Mean Reciprocal Rank (MRR).