Imbalanced metric learning for crashing fault residence prediction

As the software crash usually does great harm, locating the fault causing the crash (i.e., the crashing fault) has always been a hot research topic. As the stack trace in the crash reports usually contains abundant information related the crash, it is helpful to find the root cause of the crash. Recently, researchers extracted features of the crash, then constructed the classification model on the features to predict whether the crashing fault resides in the stack trace. This process can accelerate the debugging process and save debugging efforts. In this work, we apply a state-of-the-art metric learning method called IML to crash data for crashing fault residence prediction. This method uses Mahalanobis distance based metric learning to learn high-quality feature representation by reducing the distance between crash instances with the same label and increasing the distance between crash instances with different labels. In addition, this method designs a new loss function that includes four types of losses with different weights to cope with the class imbalanced issue of crash data. The experiments on seven open source software projects show that our IML method performs significantly better than nine sampling based and five ensemble based imbalanced learning methods in terms of three performance indicators.