Software Bug Triaging Based On Text Classification And Developer Rating

Over the past two decades,we have witnessed the growing effect of software on human society.This trend echoes the famous Marc Andreessen's 2011 proclamation that software is eating the world.Moreover,software-defined anything has been identified by Gartner as one of the top ten strategic technology trends for 2014,implying that we are entering a new era of the information society defined by software.Compared with proprietary software,open-source software(OSS)brings forth a fundamental shift in the manner in which software is developed and distributed.Due to lower cost,superior security,freedom from vendor lock-in,better quality and other primary reasons,an increasing number of enterprise-level companies have adopted OSS offerings as an essential part of their IT systems.To improve the efficiency of bug triage and reduce the cost of human and time,so far,many researchers have proposed a variety of automatic bug assignment methods.According to the data,algorithms,and tossing graph models,there are three categories of existing methods:text content-based approach,developer relationship-based approach,and hybrid approach.However,most of the existing methods assume that developers who distribute,discuss,or fix bugs are helpful to bug triaging.Therefore,they usually regard all the developers who are involved in the triaging process of a bug as the labels of the bug,so that they can get a higher recall rate when recommending possible developers.First of all,a prediction method based on text classification and developer rating is proposed in this paper.The core idea of building the prediction model is to consider both text classification based on machine learning and rating based on the source of bugs.According to the experiment on hundreds of thousands of bugs in the Eclipse and Mozilla projects,in the ten-fold incremental verification mode,the best average accuracies of our method reach 78.39%and 64.94%,respectively.Moreover,the accuracies of our method are increased by 17.34%and 10.82%,respectively,compared with the highest average accuracies of the baseline method(machine learning classification + tossing graphs).Therefore,the results indicate the effectiveness of our method.Also,we analyze the cooperation between developers based on the history of each bug report,calculate the tossing probability,and construct the tossing graph for bugs.We then combine the method mentioned above and the tossing graph to recommend potential developers related to the bug triaging process.Aiming at the defect data of the two open-source projects of Eclipse and Mozilla,the proposed approach obtains high recall rates which reach 78.20%and 71.98%,respectively.They are increased by 25.53%and 27.74%,respectively,compared with the FastText+developer rating+tossing graphs method.Also,the recall rates of our method are 14.17%and 13.25%higher than those of the ML-KNN+developer similarity method,which demonstrates the effectiveness of our method for recommending multiple possible developers.