Empirical Study: Are Complex Network Features Suitable for Cross-Version Software Defect Prediction?

Abstract

Software defect prediction can identify possible defective software modules and improve testing efficiency. Traditional software defect prediction mainly focuses on using code features and process-based features for research. The rules of complex network are suitable for software. Using complex network features to represent defect information provides a new idea for software defect prediction. In this paper, we first select 18 versions of 9 open source projects through certain rules and then build a logistic regression model based on three kinds of features (complex network features, traditional code features, merged features) to evaluate the predictive defect ability of complex network features. The results show that: (1) Compared with traditional code features, complex network features have better ability to predict defects for cross-versions software defect prediction; (2) Merged features are not as good as complex network features in defect prediction for cross-version software defect prediction, but still better than traditional code features.