The traditional collaborative filtering algorithms do not fully consider the user-item interaction information and face problems such as data sparseness or cold start, which results in inaccurate results of the recommendation system. For this reason, we propose a new recommendation algorithm, which is a collaborative filtering algorithm of graph neural network based on fusion meta-path. To be specific, first, the user-item historical interactions are embedded by a bipartite graph and the high-level features of users and items are obtained through multi-layer neural network propagation. Then, latent semantic information in the heterogeneous information network is acquired according to the random walk of meta-paths. Finally, the high-level features and latent features of users and items are combined for scoring prediction. The experimental results show that compared with the traditional recommendation algorithms, the proposed algorithm has been significantly improved.