Predicting the Severity of Open Source Bug Reports Using Unsupervised and Supervised Techniques

Abstract

The severity of the bug report helps for the bug triagers to prioritize the handling of bug reports for giving more importance to high critical bugs than less critical bugs, since the inexperienced developers and new users can make mistakes while assigning the severity. The manual labeling of severity is labor-intensive and time-consuming. In this article, both unsupervised and supervised learning algorithms are used to automate the prediction of bug report severity. Because the data was unlabeled, the Gaussian Mixture Model is used to group similar kinds of bug reports. The result is labeled data with the severity level given for each bug reports. Then, the training of classifiers is performed to predict the severity of new bug reports submitted by the user using Multinomial Naïve Bayes Classifier, Logistic Regression Classifier and Stochastic Gradient Descent Classifier. Using these methods, around 85% accuracy is obtained. More accurate predictions can be done using the authors approach.