Previous deep learning-based brain network research has made significant progress in understanding the pathophysiology of schizophrenia. However, it ignores the three-dimensional spatial characteristics of EEG signals and cannot dynamically learn the interactions between nodes. To address this issue, a schizophrenia classification model based on a three-dimensional adaptive graph convolutional neural network (3D-AGCN) is proposed. Each subject's EEG data is divided into various segment lengths and frequency bands for the experiment. The attention mechanism is then used to integrate the node features in the spatial, feature, and frequency band dimensions. The resulting adaptive brain functional network features are then constructed and fed into the GCN model. This adaptive approach eliminates the human-specified criteria for feature selection and brain network construction. The trial results demonstrated that, when using a 6-second segment length and time-domain and frequency-domain features, patients with first-episode schizophrenia achieved the highest classification accuracy of 83.11%. This method outperforms other feature selection and brain network modeling approaches, providing new insights and directions for the early diagnosis and recognition of schizophrenia.