Bug Triage Based On Data Reduction

Bug fixing is an important process in software development and maintenance.Bug triage, i.e., assigning a new bug to an appropriate fixer, is the key step of bug fixing. The main approaches to address the bug triage problem are based on text classification. However, these approaches suffer from the large-scale and low-quality data sets.In this thesis, the data reduction technique based on feature selection and instance selection is proposed to improve the accuracy of bug triage. Data reduction includes two aspects, to reduce the scale and to increase the quality. Feature selection and instance selection techniques are combined to achieve this objective. To evaluate the effectiveness of the data reduction technique, two feature selection algorithms and two instance selection algorithms are conducted on Eclipse, Gnome and NetBeans. For each data set,70%of words and50%of instances are removed. Experimental results show that the final data sets can achieve better accuracy than the original ones.From the experimental results over three project data sets, we find that the order of feature selection and instance selection has a strong impact on the final triage accuracy on different data sets. Hence, to correctly provide the best combination order for a new data set, an order prediction model is built. Continuous300000bug reports from bug repositories of Eclipse and Mozilla are chosen, and resampled to get data sets of different sizes. Each data set has18attributes. Experimental results show that the prediction model based on decision tree can achieve the accuracy of71.8%.