1 Introduction
Recommendation is attracting increasing attention in both industry and academia, owing to the prevalence and success of recommender systems in many applications [1], [2], [3], [4]. From the perspective of product providers, the aim of building a recommender system is to increase the traffic and revenue, by recommending the items that a user will be likely to consume. As such, the key data source to leverage is the past consumption histories of users, since they provide direct evidence on a user’s interest. To this end, much research effort has been devoted to collaborative filtering (CF) [5], [6], [7], which casts the task as completing the user-item consumption matrix, and incorporating side information into CF, such as textual attributes [8], [9], [10], categorical demographics [11], social information [12] and product images [13]. To utilize such diverse data in an unified model, a general class of feature-based recommendation models have been proposed, such as the pioneer work of factorization machines (FM) [14] and several recent developments that augment FM with neural networks [15], [16], [17].