Software Defect Prediction for High-Dimensional and Class-Imbalanced Data

Software quality and reliability can be improved using various techniques during the software development process. One effective method is to utilize software metrics and defect data collected during the software development life cycle and build defect predictors using data mining techniques to estimate the quality of target program modules. Such a strategy allows practitioners to intelligently allocate project resources and focus more on the potentially problematic modules. Effectiveness of a defect predictor is influenced, among other factors, by the quality of input data. Two problems which often arise in the software measurement and defect data are high dimensionality and class imbalance. This paper presents an approach for using feature selection and data sampling together to deal with the problems. Three scenarios are considered: 1) feature selection based on sampled data, and modeling based on original data; 2) feature selection based on sampled data, and modeling based on sampled data; and 3) feature selection based on original data, and modeling based on sampled data. Several software measurement data sets, obtained from the PROMISE repository, are used in the case study. The empirical results demonstrate that classification models built in scenario 1) result in significantly better performance than the models built in the other two scenarios.