Social recommendation via deep neural network-based multi-task learning

Recent years are witnessing the rapid development of social recommendation to improve the performance of recommender system, especially for cold start problem. Many of the existing social recommendation techniques utilize matrix factorization-based model, which would not essentially capture the complex nonlinear relationships for the user–item interactions and user–user interactions. Inspired by the successful application of deep neural networks on computer vision, natural language processing, and other tasks, deep learning is employed to model the social network-enhanced collaborative filtering problem. Although it has been proposed recently to model the nonlinear relationship by deep networks for collaborative filtering, it does not take the social relations between users into account. Accordingly, simultaneously modeling the social domain and item domain interactions is proposed by sharing user representation in two tasks as Social Regularized Neural Matrix Factorization (SoNeuMF). Comprehensive experiments on two real-world datasets show significant improvements of the proposed SoNeuMF framework with regard to recommendation accuracy against state-of-the-art methods.