The purpose of the recommender system is to recommend personalized products or information for users. It is widely used in many scenarios to deal with information overload problems to improve user experience. As an existing popular recommendation method, collaborative filtering usually suffers from data sparsity and cold start problems. Therefore, researchers usually make use of side information, such as contexts or item attributes, to solve the problem and improve the performance of the recommender systems. In this paper, we consider social relationship and knowledge graph as side information, and propose a multi-task feature learning model, Social-MKR, which consists of recommendation module and knowledge graph embedding (KGE) module. In recommendation module, we build the social network among users based on the user-item interactions, and conduct the GCN model to obtain the specific user's neighborhood representation, which can be used as the input of the recommendation module. Like MKR, the KGE module is used to assist recommendation module by a cross&compression unit, which can learn high-order hidden features between items and entities. Extensive experiments on real-world datasets (e.g.,  movie,book and news) demonstrate that Social-MKR outperforms several state-of-the-art methods.