Abstract: To address the issue of low recognition accuracy of collapse pillars in imbalanced seismic attribute datasets, a novel collapse pillar recognition method named "LightGBM-DNNF" was proposed, which combined Light Gradient Boosting Machine (LightGBM) and a Deep Neural Network (DNN) improved using Focal Loss. 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 binary classification and geological structure prediction. Precision, Recall, F1-score, and AUC were introduced as evaluation metrics, and comparative experiments and ablation experiments were conducted on datasets from three mining areas. Experimental results demonstrate that, compared to traditional machine learning and single ensemble learning algorithms, the F1-score and AUC values of the LightGBM-DNNFL model are both above 93%, indicating its effectiveness in recognizing collapse pillars and exhibiting stronger generalization ability.