In order to solve the problems of high sparsity of user-item interaction data and poor model interpretability in the current personalized recommendation algorithm, this paper proposes a recommendation algorithm based on lightweight neighborhood interaction graph convolutional neural network. Firstly, nonlinear Activation function and feature transformation matrix are deleted from the embedded layer of the model, the graph convolution is lighter, and the embeddings of multiple graph convolutional layers are weighted to sum to obtain the global features of the node. Second, the lightweight bilinear aggregator is used to extract the local interaction features of nodes. Third, the global features and local interaction features of the node are weighted and summed by the attention mechanism to obtain the final node embedding, the rationality and effectiveness of the proposed model were verified on three public datasets.