Abstract: To address the complexity of coal mine production accidents, the imbalance in accident categories, and the causal relationship between accident causes and accident consequences, this paper proposed a coal mine safety accident text−graph classification algorithm, CI−GAT, based on causal effects and graph attention networks. The algorithm was designed to intelligently classify text data related to coal mine safety accidents and predict the category of impending accidents based on the potential causes of accidents, and it took the CI−GNN model as its basic framework, with adaptive optimizations carried out on the GraphVAE module and the classifier module. Specifically, the algorithm first optimized the GraphVAE module: it enhanced the encoder with additional GCN layers to form a deeper GCN structure, while for the decoder part, it increased the number of MLP hidden units and incorporated BatchNorm and Dropout layers, which enabled the module to more comprehensively decode the causal nodes of accident text graphs. For the classifier module of the algorithm, the following optimizations were implemented: first, it replaced the Graph Isomorphism Network (GIN) with the Graph Attention Network (GAT) which incorporated an attention mechanism to better capture the dependency relationships between accident nodes; second, it introduced a category prototype memory to enhance minority classes and mitigate the effects of category imbalance; finally, in the multi−granularity feature fusion module, it employed a gating mechanism (FusionGate) to fuse global accident features with node features, and the fused results were then input to an MLP decoder containing two adaptive residual blocks to output the classification results of coal mine accident text. Comparative experiments and ablation experiments were conducted on a self−constructed text−graph dataset of coal mine safety accidents. The performance metrics of the proposed algorithm are as follows: accuracy is 96.3%, precision is 89.8%, recall is 93%, and F1−score is 0.913; these experimental results outperform the benchmark text−graph classification models, verifying the superiority of the proposed algorithm in classifying the text−graph dataset of coal mine safety accidents.