Applying Machine Learning Approaches to Evaluating Rockburst Liability: A Comparation of Generative and Discriminative Models

摘要

The serious consequences of rockburst have forced researchers to investigate alternatives methods for prediction. A lot of researches about rockburst resided in the focus on burst liability which is identified as an inherent cause of the rockburst. Due to the complex and highly nonlinear relationship between the impact factors and rockburst liability, traditional evaluation approaches are hard to gain ideal results for burst liability evaluation. A lot of scholars have tried to use machine learning to evaluate burst liability, but the results have been inconsistent. This study compares two fundamental machine learning models: discriminative and generative, which are typified by a support vector machine and Gaussian process classifier respectively, based on a uniform training dataset. This study also indicated burst liability evaluation is an unequal cost multi-class classification task in terms of machine learning. In addition to a conventional performance metric, the receiver operating curve (ROC) is generalized to evaluate model performances for this kind of task. The results indicate that the discriminative approach is more suitable for burst liability evaluation problem considering a common problem in burst liability evaluation task which is the sample size is limited. Finally, this conclusion was furtherly verified by a real rockburst case at a diamond mine.