Multiattribute Based Machine Learning Models for Severity Prediction in Cross Project Context

Abstract

The severity level of a reported bug is an important attribute. It describes the impact of a bug on functionality of the software. In the available literature, machine learning techniques based prediction models have been proposed to assess the severity level of a bug. These prediction models have been developed by using summary of a reported bug i.e. the description of a bug reported by a user. This work has been also extended in cross project context to help the projects whose historical data is not available. Till now, the literature reveals that bug triager assess the severity level based on only the summary report of a bug but we feel that the severity level of a bug may change its value during the course of fixing and moreover, the severity level is not only characterized by the summary of bug report but also by other attributes namely priority, number of comments, number of dependents, number of duplicates, complexity, summary weight and cc list. In this paper, we have developed prediction models for determining the severity level of a reported bug based on these attributes in cross project context. For empirical validation, we considered 15,859 bug reports of Firefox, Thunderbird, Seamonkey, Boot2Gecko, Add-on SDK, Bugzilla, Webtools and addons.mozilla.org products of Mozilla open source project to develop the classification models based on Support Vector Machine (SVM), Naïve Bayes (NB) and K-Nearest Neighbors (KNN).