(An) effective software defect prediction via source code to image conversion with semantic information

Abstract

Primary goals of Software Defect Prediction (SDP) are to predict potential defect-prone source codes in a software and help developers to easily find defects, effectively allocate valuable resources for software quality assurance. The two major steps of SDP are feature extraction and training the model, and many SDP studies have focused on the feature extraction phase. Traditional SDP studies have used manually designed features, which include Halstead metric, McCabe metric, and the change history metric. However, these features depend on how the code is written and changed, thus they do not consider the semantic information of the code. Since semantic information contains how the code works, this information is essential to analyze the defectiveness of the source code. Recent SDP studies proposed the technique that extracts the feature with semantic information. Some studies used Abstract Syntax Tree (AST) to convert whole source code into one vector representing the feature, and trained the defect prediction model using various methods, such as Deep Belief Network and Convolutional Neural Network. Other study used AST to convert each AST node into vector, analyzing the defectiveness of the source code with finer granularity. Finally, a technique which leveraged powerful image classification network and converted the source code into image to classify the defect pattern in the image showed the higher accuracy of the prediction. However, these studies did not fully utilize the semantic information of the source code, which can degrade the prediction performance. In this thesis, we are motivated by the SDP study that uses a powerful image classification network, and we suggest the technique that preserves the semantic information when converting to an image. In order to preserve the semantic information, we extract each phrase (keywords, variables, operators) in the source code, and map to a unique color. We conducted experiments on 8 projects for comparing the accuracy with existing SDP techniques. The results show that our approach is more accurate than the existing techniques.