Information gain based dynamic support set construction for cold-start recommendation

A fundamental challenge for recommendation systems is the cold-start problem, i.e., recommending with no or few user-item interactions. An emerging direction alleviates the problem with meta-learning. These methods utilize the prior knowledge learned from previous recommendation tasks to facilitate the learning of a new task with a small number of interactions, i.e., the support set. The new task involves making recommendations for cold users or items. However, a support set may contain uninformative items and fail to capture the personality of new users, jeopardizing the recommendation quality. This paper proposes Information gain driven cold-start Recommendation (InfoRec), which actively acquires informative items for constructing the support set. Specifically, our InfoRec quantifies item informativeness with the reduction of model uncertainty about the user preference for items that are not in the support set. Experiments demonstrate that InfoRec outperforms baseline methods on two benchmark datasets in both non-cold-start and cold-start scenarios.