Learning Multi-aspect Shared Representation for Recommendation

Representation learning-based recommendation models have garnered considerable attention for their ability to extract meaningful patterns from historical user–item interactions. Nevertheless, most existing approaches represent each user or item with a single embedding, which limits their ability to capture the multidimensional characteristics of users and the multivariate attributes of items, and often underutilizes shared traits across similar users or items. To address these limitations, we propose LMR-RS, a framework for learning multi-aspect shared representations for recommendation. LMR-RS maintains two shared representation pools for users and items, from which multiple relevant embeddings are dynamically selected to construct a multi-aspect representation for each entity. A scaled dot-product attention mechanism is further employed to model fine-grained alignment between user aspects and item aspects when computing preference scores. To further enhance representation quality, we introduce a diversity loss to discourage redundancy among pool embeddings and a consistency loss to align the selected embeddings with the underlying user and item representations. Experiments conducted on six real-world datasets from different domains demonstrate that LMR-RS achieves competitive and robust performance in terms of Recall@20 and NDCG@20, while comprehensive ablation studies further confirm the contribution of each model component.