Graph convolutional network (GCN) is a deep learning model for graph signal processing and has been used in many real-world applications due to its powerful ability of feature extraction. As the recommendation problem can be viewed as link prediction of graph signals, recently several GCN based methods have been proposed for recommender systems. A recommender system involves two kinds of interactions, with one representing interactions between users and items and the other representing interactions among users (or items). However, existing methods focus on either heterogeneous or homogeneous interactions only, thus their modeling expressiveness is limited. In this study, a new GCN based recommendation algorithm is proposed to jointly utilize these two types of interactions. Specifically, a heterogeneous convolutional operator is used to mine information from the spectrum of user-item graphs, while a homogeneous convolutional operator is used to enforce similar vertices to be similar in the hidden space. Finally, the experiments on benchmark datasets show that the proposed method achieves better performance compared with several state-of-the-art methods.