Comparative Analysis of Prediction Performance of Mine Subsidence Using Machine Learning Techniques

In this study, the prediction of mining-induced subsidence is analyzed and compared using various machine learning models. Factors affecting the occurrence of subsidence are identified from eight and 1,730 sets of subsidence data. Five machine learning models are selected, i.e., Adaboost, artificial neural networks, the k-nearest neighbor, random forest, and the support vector machine, which are frequently used in studies related to geohazard prediction. In addition, the stacking technique is applied to five algorithms based on 10 combinations, and the predictive performance of each ensemble method is evaluated and compared. To evaluate the classification performance of the machine learning technique applied in this study, recall is used as an evaluation index, which describes the ratio of the predicted ground subsidence instead of the area under curve used previously. Based on the values of recall, the random forest demonstrates the best performance (with a recall of 0.955). The recall is expected to be a more reliable evaluation index for predicting subsidence occurrences compared with other indices.