Abstract: In order to solve the problem of low recognition accuracy of collapse column in unbalanced seismic attribute datasets, a novel collapse column recognition method named "LightGBM-DNNF" was proposed, which based on the light gradient boosting machine (LightGBM) and the use of focal loss to improve deep neural networks (DNN). Firstly, attribute selection was performed through correlation and importance analysis. Then, the paths of LightGBM leaf nodes were extracted as new features, which were combined with the original dataset to form a new dataset. Finally, the new dataset was input into the DNNFL model for classification training and prediction of geological structure types. Precision (P), Recall (R), F1-score (F_1-score), and area under the curve (A_UC) were introduced as evaluation metrics, and comparative experiments and ablation experiments were conducted based on datasets from three mining areas. The experimental results show that compared with the traditional machine learning and single ensemble learning algorithm, the F_1-score and A_UC values of the LightBM-DNNFL model are both above 93%, indicating its effectiveness in recognizing collapse column and exhibiting stronger generalization ability.