The friend recommendation service is instrumental in molding and enhancing the expansion of online social networks. Graph Neural Networks (GNNs) have become a powerful tool for learning graph data and have been widely used in the design of recommender systems, as they can effectively aggregate the neighborhood representation of nodes. The basic idea of GNNs-based recommender systems is to utilize GNNs to aggregate a user’s neighborhood information in the user-user graph and the user-item graph. For friend recommendation, combining users’ neighborhood information in the user-user graph and the user-item graph makes recommendations more diverse. However, one of the challenges is how to effectively combine the neighborhood information of users from the two graphs. In addition, since users usually have many items to interact with, accurately analyzing the user’s personal preferences is a challenging task. In this paper, we present a GNNs-based framework (FriendRec) for friend recommendations. We first utilize GNNs to aggregate social representation and interest representation, respectively. Then we fuse two kinds of representations with a different strategy from previous works. In particular, we design a special layer to aggregate interest representation, which achieves 5% uplift compared with state-of-the-art recommendation model. Compared to fusing representations with deep neural networks, FriendRec significantly reduces the number of parameters, making it easier to train.