Abstract: Traditional stock price prediction methods usually treat it as a time series forecasting problem, without utilizing the spatial information of the stocks. On the other hand, using Graph Neural Networks (GNN) for prediction lacks the graph data structure of the relationships between stocks. This paper proposes a unidirectional adaptive Graph Neural Network model based on the similarity of trends between stocks within the same sector and the significant impact of leading stocks on other stocks, called Stock-GNN, to predict stock prices. The unidirectional graph learning module of the model can adaptively train the graph data structure between stocks, and then extract features in time and space using the temporal convolution module and the graph convolution module. The real data of the four major sectors of A shares (banks, liquor, power equipment, and biomedicine) were crawled for empirical analysis, and compared with existing stock prediction models such as AR, GARCH, CNN-GRU, TPA-LSTM, CNN-LSTM-Attention, and an all-sector Graph Neural Network (All-SGNN). The experimental results show that the Stock-GNN prediction model has lower prediction error and higher correlation coefficient.