Recommender systems have been widely adopted in various large-scale Web applications. Among these applications, online dating application has attracted more and more research efforts. Essentially, online dating data is a bipartite graph with sparse reciprocal links. Reciprocal recommendations consider bi-directional interests of service and recommended users, not merely the service user’s interest. This paper proposes a motif-based graph neural network (MotifGNN) for online dating recommendation task. We first define seven kinds of motifs and then design a motif based random walk algorithm to sample neighbor users to learn feature embeddings of each service user. At last, these learned feature embeddings are used to predict whether a reciprocal link exists or not. Experiments are evaluated on two real-world online dating datasets. The promising results demonstrate the superiority of the proposed approach against a number of state-of-the-art approaches.