Abstract Software maintenance is a challenging and laborious software management activity, especially for open‐source software. The bugs reports of such software allow tracking maintenance activities and were used in several empirical studies to better predict the bug resolution effort. These reports are known for their large size and contain nonrelevant instances that need to be preprocessed to be suitable for use. To this end, instance selection (IS) has been proposed in the literature as a way to reduce the size of the datasets, while keeping the relevant instances. The objective of this study is to perform an empirical study that investigates the impact of data preprocessing through IS on the performance of bug resolution prediction classifiers. To deal with this, four IS algorithms, namely, edited nearest neighbor (ENN), repeated ENN, all‐k nearest neighbors, and model class selection, are applied on five large datasets, together with five machine learning techniques. Overall, 125 experiments were performed and compared. The findings of this study highlight the positive impact of IS in providing better estimates for bug resolution prediction classifiers, in particular using repeated ENN and ENN algorithms.

In software engineering, maintenance effort estimation is a challenging research topic. Several empirical studies have focused on Maintenance Effort Estimation for Open Source Software (O-MEE) using Machine Learning (ML) techniques. Although Tuning the Parameters (TP) of ML techniques has proved to improve their performance, it has not yet been investigated in the context of O-MEE for bug resolution prediction. Therefore, this empirical study investigates the impact of grid search TP method, on the performance of k-Nearest Neighbor, Support Vector Machine, and Random Tree using Eclipse Platform, Eclipse JDT, and Apache datasets. A set of 18 ML classifiers were built and the results show the positive impact of the TP on the ML techniques' performance compared to the default ones.

Studies about maintenance effort estimation of open-source software have investigated the impact of single instance selection on the performance machine learning techniques. However, Ensemble of Instance Selection (EIS) has not been investigated for this topic, especially for bug resolution prediction. This empirical study considers the impact of EIS on the performance of k-Nearest Neighbor, Support Vector Machine, and Multinomial Naïve Bayes techniques. A set of 27 classifiers are built using Bagging and Random Feature Subset Ensembles based on AllkNN single instance selection algorithm on three datasets. The results are presented, together with a comparative analysis of the built classifiers' performance. The results show that the investigated EIS algorithms improve the performance of ML classifiers and outperform the single instance selection algorithm-based classifiers.