Bug Severity Prediction And Bug Triage Model Based On Neural Network

With the increasing of scale and complexity of the software,the softwares' defects become unaviodable,which leads to software maintenance becoming quite important.Therefore bug fixing is an essential part of software maintenance.The primary task in bug fixing is to manage bug reports.Currently,there are a number of research sites that manage bug reports,in which,Bugzilla are widely used in the open source.When a bug report is submitted,the website manager will determine the severity level of the bug,thereby according to the severity level of this bug,judge emergency of this bug,and establish reasonable fixing scheme which can save time and resource.So bug severity prediction is an important prerequisite for bug fixing.After judging the severity of the bug,the website manager will allocate appropriate developer to fix the bug.If an inappropriate person is allocated,it will be allocated again,which will lengthen the bug fixing cycle and increase labor costs.So it's also important to find the right fixer quickly.Now,most of the work described above is done manually,resulting in a huge waste of developer and a bug fix heap lag.In this paper,we proposed bug severity prediction model and bug fixer triage model,which are both based on neural network.These two model can automatically predict the severity of bug reports,and then allocate appropriate developer to fix.In our method,first,we no more use the bug description as raw data like the past research.Instead,we consider the multi-dimensional data,such as components,the submitter,hardware platform,which are different dimensions,and different types.In addition,the modified GRU recurrent neural network model is used in the bug prediction model,and it is noted that the word order can be taken into account in the process of text processing.In the bug fixer triage model,we not only consider the bi-GRU,but also consider that the convolutional neural network can automatically extrac features of the text.Combinng the two models' advantages,and several improvements have been made to the structure of a single model,result in a new CNN-RNN model.In order to verify the methods proposed in this paper,we measured the effectiveness of our study by using a total of 180,000 bug reports extracted from five open products from the bugzilla website(platform,cdt,jdt,pde and birt).The experiment results demonstrate that our method achieves a higher accuracy and recall rate(0.71~0.82)when predicting the severity of the bug.When recommanding bugfixer,top5 accuracy was more than 65%,and top10 accuracy reached more than 80%.The above data proves that our method is superior to the previous bayesian and theme-based models.Our research provides a reference for researchers in this field.