Predicting Bug-Fixing Time with Rating Features - A Comparison Between Ratings and Reputations

Abstract

Predicting bug-fixing time plays an important role in allowing asoftware manager and team to make decisions about allocationof resources, prioritization and scheduling. Estimating the timeto fix a bug is not a simple task. In the literature, machine learning(ML) models have been proposed to help software managersdecide whether a bug might be fixed now or later. One featurehighlighted in ML models for predicting bug-fixing time is reporterreputation. However, these features are based on the participationof the reporter or developer in the project, but do not take intoaccount the time taken to fix the bugs. In this study, we proposenew two features called "reporter rating" and "developer rating."Unlike reputations, ratings are based on the time taken to fix abug. In this study, we carried out an experiment in two datasetscontaining bug reports.We ran the reputation and rating features inten ML models and compared the results. Additionally, we verifiedthe features together and combined them with textual features. Asa result, we found that ratings can improve the performance of themodels. Ratings had the best results in probabilistic models, whilereputation was better in models that use the decision tree approach.When used together, reputations and ratings do not substantiallyincrease the performance of the models when compared to individualresults. However, ratings improve performance when combinedwith textual features more than reputations.