Abstract: To address the issue that the existing multi-behavior recommendation systems do not effectively use di-fferent level graphs to spread information, and it is difficult to capture the impact strength of different actions on user decision making, a multi-behavior recommendation model based on graph attention network (GABR) is proposed. First, a mini-batch sampling mean aggregation method is used to aggregate neighboring nodes interacting with the same behavior type to improve the feature representation efficiency. Next, an attention mechanism is employed to learn the influence coefficients of different behavior types to further fuse node features. Then, multi-layer user and item representations are merged to effectively utilize graphs at different levels to propagate information. In the end, in order to optimize the model performance, the user-item pairs with interactions and randomly sampled user-item pairs without interactions are used as positive and negative sample pairs to train the target model.To verify the model recommendation performance, the GARB model is compared with 9 existing recommendation models on three real datasets (Yelp, Scholat, Beibei). The experimental results show that the GABR model can effectively use multi-layer graphs with influence coefficients of different behavior types to better predict user preferences. Compared with the best baseline model (GNMR) on three real datasets, the HR and NDCG of the GABR model increase by 1.73% and 2.43% on average.