Handling new users and items: a comparative study of inductive recommenders

Abstract

Usually, recommender systems are trained on a set of users and items and then used to recommend new user-item pairings among those seen during training. As users and items are added continuously, there is a pressing need to provide recommendations for new users and items, i.e., for users and items not seen during training. Solutions to this problem exploit techniques like meta-learning or auxiliary information encoded in knowledge graphs to learn an “inductive bias”. Yet, most existing works can either recommend for new users or new items not seen during training but not both. Further, existing methods have rarely been compared to each other. Finally, existing evaluations of these methods use a random split of training data, and thus do not consider temporal splits of ratings in training and testing. This setting ensures testing is correctly performed on user interactions that actually occur after the training period. In this paper, we propose a framework for training and testing the methods on three real world datasets, and perform a deeper analysis of each dataset to better understand the effect of emerging popularity trends. As a result, our re-evaluation of state-of-the-art methods identifies strong architectures and solutions for inductive recommendation. We find that inductive methods that perform aggregation are able to outperform non-aggregating methods in all settings; performances vary greatly across settings, pointing to new important research questions.