Cost-Driven Automatic Bug Triage

Bug triaging is the process of assigning bugs to corresponding bug fixers.It is an important function in bug management system.Bug triaging is mostly done manually at present.However,software’s high complexity and assigners’ lack of essential knowledge makes bug triaging a time-consuming work with low accuracy.Moreover,people always hope to find a reasonable triaging strategy which assigns bugs quickly and improves the triaging result while decreasing the human labor.Through rigorous analysis,it is found out that bug fixing time is an important criterion for evaluating the triaging result.However,existing approaches typically treat bug triaging as a problem of accuracy optimization,which leads to high-cost triaging result and lengthens the fixing time of bugs.In some applications,traditional method which only considers accuracy is not even applicable.Thus to avoid triaging results with high time cost,the proposed framework considers the time cost of a triaging scheme.It aims at reducing the time cost without much influence to the accuracy.To balance the influence of accuracy and time cost,investment portfolio model is used to help produce the final triaging result.This framework is also designed to produce different triaging scheme according to different scenarios.The procedure of our framework is as follows.First,we turn bug reports into word bags and form two vectors-word count vector and topic probability vector;in offline learning stage,the weight between these two vectors is learnt by performing experiments on the validation set.With this weight,the above two vectors are merged into one,which is used to compute the similarity between bug reports;in the online triaging stage,investment portfolio model is used.By altering the value of its hyper-parameter,this model produces investment schemes with different emphasize on investment risk and expected yield.To adapt the economic model to this problem,the risk and yield in the model are changed to accuracy and time cost respectively.And the hyper-parameter is taken as the importance of time cost;in the end,the similarity vector and the time cost vector is used as the input of the refined model to produce the final triaging result.According to the experimental results,we confirm the necessity of taking time cost as an optimization goal;by varying the value of the hyper-parameter,the triaging results tend to have different emphasis on accuracy and time cost;In contrast with related method-Costriage,the experimental results show that our model produces a lower-cost triaging scheme with the same accuracy,and it also reduces the time needed to train the model.