In the recommender system, compared with the traditional neural network, the neural network based on the knowledge graph can combine node information and topological structure to infer and recommend with grpahs as input. However, the existing recommender algorithm based on graph neural networks faces the problem of inaccurate knowledge representation and single information fusion. Combining graph neural network with attention, the bias-based graph attention neural network recommender (BGANR)algorithm is proposed. First, the translation model is used in the feature representation to gain the triple information of the node in the same projection space. Considering the error between the predicted value and the true value in the triple, and the difference in the weight of neighboring nodes during information dissemination, the bias-based attention method is used to better capture the high-order connectivity of nodes. Then, in the propagation training of the neural network, the node and neighbor information is aggregated through the multi-channel fusion mechanism to increase the robustness of the model. Finally, comparison with the state-of-the-art algorithms on two real datasets verifies the effectiveness of the BGANR algorithm.