Predicting bug severity in open-source software systems using scalable machine learning techniques

Abstract

As software systems become larger and more complicated, the task of detecting and fixing bugs to improve the software performance is getting more tedious and inefficient. Automated processes that detect and report bugs quickly and with high accuracy are needed. In this thesis, we describe an approach, which is fast and performs the bug classification task with comparatively better accuracy then previously reported research. Here, we used the machine learning methods, specifically an online algorithm for bug classification. This approach involves the use of text mining algorithm for feature extraction. Then the data is used to train classifier models using an online machine learning classification algorithm for optimized performance. The above steps are done twice, once for a binary model and once with a multi-class model. The multi-class model predicts as many as seven bug severity levels with the aim of prioritizing the bug assignment process. After analyzing all four datasets collected from open source software system, we can predict good with accuracy (72%-98%) if the data is balanced and has sufficient size of training set.