Recommender systems leverage both content and user interactions to generaterecommendations that fit users' preferences. The recent surge of interest indeep learning presents new opportunities for exploiting these two sources ofinformation. To recommend items we propose to first learn a user-independenthigh-dimensional semantic space in which items are positioned according totheir substitutability, and then learn a user-specific transformation functionto transform this space into a ranking according to the user's pastpreferences. An advantage of the proposed architecture is that it can be usedto effectively recommend items using either content that describes the items oruser-item ratings. We show that this approach significantly outperformsstate-of-the-art recommender systems on the MovieLens 1M dataset.
展开▼
