Recommendation in feature space sphere

Recently, recommendation algorithms have been widely used in many e-commerce platforms to recommend items to users on the basis of their preferences to improve selling efficiency. Matrix factorization methods which extract latent features of users and items by decomposing the rating matrix have achieved success in rating prediction. But almost all of these algorithms are designed to fit the rating matrix directly to get the latent features and ignore the user-item relationship in feature space. To this end, in this paper, we propose a recommendation in feature space sphere (RFSS) which takes into account the relationship between users and items in feature space. Different from the conventional latent feature based recommendation algorithms, the proposed algorithm supposes that if a user likes an item, the user is close to the item in feature space. Meanwhile, the closer a user and an item are in feature space, the higher the predicted rating will be. And an adaptive user-dependent coefficient is introduced to map the user-item distances to the predicted ratings. Extensive experiments on four real-world datasets have been conducted, the results of which show that our proposed method outperforms the state-of-the-art recommendation algorithms.