Research On Software Defect Prediction Based On Program Slice

Defects are inherent problems in software and can cause a lot of unexpected troubles to the system and users.It has become an indispensable part in software production to predict software defects and correct them.Most research of the current software defect prediction technologies mainly focus on the design of feature related to the potential defect code and the combination of tree representation of program and deep learning.However,these methods cannot fully express the syntax and semantic information of the source code.This paper proposes a software defect prediction model based on program slice.The model can analyze control flow and data flow through the program slice based on system dependency graph.The method of program slice can preserve the semantic information between program statements.This paper designs a deep learning model that combines word vector model with Gated Recurrent Unit neural network.The word vector model is then used to transform the program slice into a fixed-length vector representation.The vectors of the training set and their labels(whether they contain defects)are used to train the neural network model.Finally,the trained neural network model is used to predict the software defect of the test set.A series of experiments performed on ten open source projects of the PROMISE dataset show that compared to existing defect prediction methods(AST-DBN,tree-LSTM and Seml),the method proposed in this paper outperforms both within-project defect prediction(WPDP)and cross-project defect prediction(CPDP).Specifically,for the defect prediction within the project,in terms of average F1 value,the method in this paper has improved by 4.2%,7.1% and 1.5% respectively compared with AST-DBN,tree-LSTM and Seml.For cross-project defect prediction,in terms of average F1 value,the method in this paper is better than ASTDBN,tree-LSTM and Seml by 3.9%,7.2% and 4.1%.Through the experiment verifies the effectiveness of the method in this paper.