Empirical validation of feature selection techniques for cross-project defect prediction

In software engineering, cross-project defect prediction is an important area in the field of defect prediction and change prediction. Software defect prediction aims for the identification of defects in the early stages of the software development life cycle. Defect prediction helps in optimizing the resources in terms of testing and reduction of maintenance efforts. Defect prediction works appropriately if a huge amount of data is available for training a prediction model. Nowadays, a sufficient amount of data is not available due to which we have to use different data for training and testing of a project. The cross-project defect prediction idea emerged when different projects are used as training and testing dataset for the identification of defects. To design a defect prediction model we have to consider only significant features out of all the features set. Feature selection techniques are categorized based on unsupervised and supervised learning. The major limitation of cross-project defect prediction is handling different data distributions of source and target projects. The experiment was conducted using AEEEM and ReLink software defect dataset. Moreover, five projects of AEEEM and three projects of ReLink with a maximum count of files in the selected projects are 1862 and 399. In this study, we have analyzed the effect of feature selection techniques in cross-project defect prediction. The results were analyzed using AUC. The significance of filter, wrapper, and swarm search-based methods for feature selection techniques was analyzed separately. There is a trade-off between computational complexity and the performance of feature selection techniques. Swarm search-based methods performed better than filter and wrapper methods in terms of computational cost and overall performance of the prediction modes. The results were statistically validated using Friedman test and Wilcoxon signed rank test.