Bug Severity Assessment in Cross Project Context and Identifying Training Candidates

Abstract

The automatic bug severity prediction will be useful in prioritising the development efforts, allocating resources and bug fixer. It needs historical data on which classifiers can be trained. In the absence of such historical data cross project prediction provides a good solution. In this paper, our objective is to automate the bug severity prediction by using a bug metric summary and to identify best training candidates in cross project context. The text mining technique has been used to extract the summary terms and trained the classifiers using these terms. About 63 training candidates have been designed by combining seven datasets of Eclipse projects to develop the severity prediction models. To deal with the imbalance bug data problem, we employed two approaches of ensemble by using two operators available in RapidMiner: Vote and Bagging. Results show that k-Nearest Neighbour (k-NN) performance is better than the Support Vector Machine (SVM) performance. Naive Bayes f-measure performance is poor, i.e. below 34.25%. In case of k-NN, developing training candidates by combining more than one training datasets helps in improving the performances (f-measure and accuracy). The two ensemble approaches have improved the f-measure performance up to 5% and 10% respectively for the severity levels having less number of bug reports in comparison of major severity level. We have further motivated the paper with a cross project bug severity prediction between Eclipse and Mozilla products. Results show that Mozilla products can be used to build reliable prediction models for Eclipse products and vice versa in case of SVM and k-NN classifiers.