SphereCF: Sphere Embedding for Collaborative Filtering

Recently, metric learning has shown its advantage in Collaborative Filtering (CF). Most works capture data relationships based on the measure of Euclidean distance. However, in the high-dimensional space of the data embedding, the directionality between data, that is, the angular relationship, also reflects the important relevance among data. In this paper, we propose Sphere Embedding for Collaborative Filtering (SphereCF) which learns the relationship between cosine metric learning and collaborative filtering. SphereCF maps the user and item latent vectors into the hypersphere manifold and predicts by learning the cosine similarity between the user and item latent vector. At the same time, we propose a hybrid loss that combines triplet loss and logistic loss. The triplet loss makes the inter-class angle between the positive and negative samples of a user as large as possible, and the logistic loss makes the intra-class angle of the positive user-item pairs as small as possible. We consider the user and item latent vector as a point on the hypersphere, which makes the margin in the triplet loss only depend on the angle, thus improving the performance of the model. Extensive experiments show that our model significantly outperforms state-of-the-art methods on four real-world datasets.