Instance Selection for Online Updating in Dynamic Recommender Environments

Online recommender systems continuously learn from user interactions that occur in a streaming manner. A fundamental challenge of online recommendation is to select important instances (i.e., user interactions) for model updates to achieve higher prediction accuracy while omitting noisy instances. In this paper, we study (1) how to select the best instances and (2) how to effectively utilize the selected instances in dynamic recommender environments. We present two instance selection strategies based on Self-Paced Learning and rating profiles. We integrate them with Factorization Machines to perform online updates. Moreover, we study the impact of contextual information in online updating. We conducted experiments on a real-world check-in dataset, which contains temporal contextual features. Empirical results demonstrate that ox ur instance selection strategies effectively balance the trade-off between prediction accuracy and efficiency.