Community-Level Personalized Recommendation by Exploiting Evolving User-Item Micro-Clusters

To achieve precise personalized recommendations, it is crucial to model user preference and item popularity simultaneously to capture their evolving characteristics. However, existing techniques often treat the evolution of user preference or item popularity with the same fixed decay rate, failing to capture the individual evolution pattern of each user or item. Beyond, the co-evolution patterns of users and items are not well learned due to the data sparsity of each user-item interaction. In this paper, we propose a new personalized recommendation algorithm, called EvoRec, by exploiting a set of dynamic user-item micro-clusters and modeling the evolving co-evolution patterns at the community level. Specifically, building upon temporal graph learning, each user or item embedding is represented as an evolving microcluster, and incrementally updated with a personalized decay factor. Afterwards, dynamic community structure of microclusters is uncovered to exploit the user-item co-evolution pattern, and finally support robust recommendation. Compared with established methods, EvoRec has several advantages: (a) It employs a unique adaptive learning mechanism that captures evolving dynamics of each user and item individually. (b) By exploring the user-item co-evolving patterns at the community level, EvoRec not only allows capturing intricate relationships between user-item interactions, but also support robust recommendation. Our experiments on real-world datasets indicate that EvoRec outperforms state-of-the-art baselines, and has the capacity to promptly capture shifts in user preferences and item features.