Classificação de issues obtidas de repositórios de software: uma abordagem baseada em características textuais

Abstract

The classification of issues in software maintenance repositories is currently done by software developers. However, this classification is conducted manually and is not free of errors, which cause problems in the distribution of issues to the maintenance teams. This happen because the developers, which usually are the proponents of the issues, have the bad habit of classifying them as bugs. This erroneous rating generates rework and other disadvantages to the teams. Therefore, the main objective of this study is to improve this classification, using an issue classification approach conducted in an automated manner. In turn, this approach was implemented based on machine learning tecniques. These tecniques show that keywords discriminant of issues types can be used as attributes of automatic classifiers for prediction of these issues. The approach was evaluated on five open source projects extracted from two widely used issue trackers, Jira and Bugzilla. Because they are old issue trackers, the chosen projects provided good number of issues for this study. These issues, about 7.000, were classified by human experts at work [Herzig, Just e Zeller 2013], producing a feedback which was used for this study. This present work produced an automatic issues classifier, with 81% of accuracy, able to predict them in types of bug, request for enhancement and improvement. The result of accuracy obtained by this classifier suggests that it can be used in delivery systems to treatment teams with the purpose of reducing rework that occurs in these teams because of the poor issues rating.