Towards Non-Linear Social Recommendation Using Gaussian Process

Abstract

Recent research on recommender systems has proved that by leveraging social network information, the quality of recommendations can be evidently improved. Traditional social recommendation models typically linearly combine social network information. For instance, matrix factorization based models linearly combine latent factors of relevant users and items. However, in practice, the multifaceted social relations are so complex that simple linear combination may not be able to reasonably organize such information for accurate social recommendation. On the other hand, existing deep learning based non-linear methods lack systematic modeling of user-item-friend relations. To handle these issues, we propose a novel, non-linear latent factor model for social recommendations leveraging Gaussian process. By introducing a social-aware covariance function, we organize individual users' past feedback, as well as the associated social information (e.g., friends' feedback to the same items) into a covariance matrix, which non-linearly and systematically learns the complex interactions among users, their interacted items and their friends' opinions. A stochastic gradient descent based optimization algorithm is developed to fit the model. Extensive experiments conducted on three real-world datasets demonstrate that the proposed model outperforms the state-of-the-art social recommendation models and Gaussian process based models.