Learning Popularity-Aware User Representations for Sequential Recommendation via Retrieval-Augmentation

Sequential recommendation methods are important in recommender systems, as they capture users’ dynamic interests from their historical behavior sequences. However, most existing methods have two limitations. First, they heavily rely on semantics from only a target user's interaction sequence, despite users and items with limited interactions being prevalent in real-world scenarios. Second, they suffer from popularity bias. Although data augmentation techniques have been proposed to address these issues, they struggle to augment the user with relevant and diverse data, and have limitations in dealing with popularity bias or responding to it excessively. Hence, we propose a new sequential recommendation framework that learns popularity-aware user representations via retrieval-augmentation, called REDS, to improve the quality of recommendations for users with sparse interactions and to mitigate popularity bias by: (i) incorporating embeddings of how often each item occurs in all users’ interactions into a user representation, (ii) augmenting the representation with her/his top- \(k\) similar users’ representations by leveraging the user's supervision signals during training, and (iii) disentangling the user preference into two factors to capture complementary semantics between the target user's and her/his similar users’ interaction sequences. Experiments on three real-world datasets show that REDS outperforms state-of-the-art baselines. The source code is available at https://github.com/JaneHubb/REDS .