Research Of Neural Network Based Bug Triage And Data Misuse

To manage software bugs conveniently,many large open-source projects use the bug tracking system to records and manages bugs in terms of bug reports.Bug triage is an important step in bug management,which aims to assign a suitable developer to fix a bug.The efficiency of fixing bugs is critical to software quality,which requires assigning developers to bugs quickly and accurately.However,in large open-source projects,not only the amount of daily new bugs is a lot,but also the number of developers that exist in the projects is large.So bug triage consumes many manual costs and time costs.To improve bug triage efficiency,previous researches propose many methods to predict the fixers of bugs.The newest kind of method is using the neural network to train a classifier.But the datasets and hyper-parameters used in these methods are different,which makes it is hard to compare these methods.So we construct similar models according to the methods and conduct experiments in the same datasets and similar hyper-parameters to compare the models more easily.Moreover,we use the attention mechanism and label smoothing to enhance existing models.Lastly,because the attributes of bugs can change over time,it could get too optimized results if using attributes of the wrong time.So we use attributes of different time to explore the effect of data misuse on model performance.The main work of this paper as follows: 1.Using basic attributes of bugs like text and component as features,using convolution neural network,recurrent neural network or attention mechanism to construct models.2.Using derived attributes of bugs like within component fixer sequences as the feature,explore the effect of the form of fixer sequences and dynamic label smoothing according to fixer sequences.3.Exploring the effect of adding bug reporter attribute as the feature on model performance in total and in different kinds of bug reports.4.Exploring the effect of data misuse on model performance.The main conclusions of this paper as follows: 1.Features are far more important than the type of neural network on model performance.Only use derived attributes can get a better model than models using basic attributes.2.Dynamic label smoothing can improve the accuracy of the later part in the predicted fixers list.Static label smoothing can improve the accuracy of the former part in the predicted fixers list.3.The improvements in model performance when adding bug reporter as the feature almost come from the part of bugs whose reporters equal to their fixers.The value of improvements is limited because reporters can assign the bug to themselves in this situation.4.Each attribute change has an obvious effect on model performance,the order of effects from large to small are fixer sequences,component,and text.