Fixer Recommendation Of Software Bugs With Multi-source Data

The software defect repair(also known as bug fixing)process records the life cycle of a software defect,including a series of states such as NEW,CONFIRMED,ASSIGNED,FIXED,VERIFIED,and CLOSED.During this process,it involves many ways of cooperation between developers,such as communicating bug information by emails and assigning bugs to different developers,which form multi-source data in the process of defect repair.At present,most open-source software uses a bug tracking system to record and manage defects to repair them efficiently.With the continuous expansion of the scale and the explosive growth of the number of open-source software,a large number of defects have been reported to the bug tracking system.In order to achieve the purpose of efficient defect repair,in recent years,many researchers have studied automated bug triaging approaches.However,the existing methods have some limitations in predicting bug fixers.First,developer collaborative network-based recommendation methods only consider the structural measures of single-layer graphs and ignore the multiplex(or multi-layer)network composed of different types of developer behaviors.Second,features extracted from text or structure cannot be combined effectively,resulting in weak recommendation effects and lacking scalability.For the first problem,this study attempts to analyze multi-source data generated during the process of defect repair.We integrate such multi-source data into a developer collaborative multiplex network,expand the measures of single-layer networks to those of the multiplex network,and obtain the structural vectors through network embedding.Besides,the text features of defects are obtained through the LDA(Latent Dirichlet Allocation)model,and then directly combined with the structural features to train the recommendation model.The experimental result shows that the multiplex network embedding can learn the network structure of developer collaboration better.Using the multi-layer perceptron model,the Top-1 prediction accuracies on the two data sets are 0.614 and 0.586,respectively,which are 2.9% and 4.3% higher than the best result of the benchmark method.Considering the combination of different types of data(network structure and bug text data)for the second problem,we use GCN(Graph Convolutional Network)as a joint classifier to multiply the text features extracted by Text CNN.The developer ID corresponding to the maximum value of the product is used as the recommended fixer.Experiments show that using GCN as a joint classifier multiplied with text embedding feature can get better prediction results.Compared with the results obtained for the first problem,the Top-1 prediction accuracies on the two data sets are further improved,which are 18.6% and 9.7% higher than those of the classifier in the first problem.