An Ensemble Feature Selection Method Based on Symmetrical Uncertainty and Correlation for Software Defect Prediction

Discovering and fixing potential defects in the software before it is put into use can ensure the reliability of the software. Software defect prediction technology enables the early detection and identification of software defects, thereby offering valuable insights for software testers to allocate their testing resources in a rational manner. This technology can enhance the efficiency and effectiveness of the testing process. However, the excessive dimension of features within software datasets poses a challenge, limiting both the efficiency and effectiveness of software defect prediction models. Thus, reducing the dimension of software datasets by eliminating irrelevant and redundant features that are not critical for defect prediction is of great significance in improving the performance of software defect prediction model. Comparative experiment has been implemented to evaluate and contrast the comprehensive application performance of seven classical feature selection algorithms when applied to diverse software datasets. This study introduces the principal component and proposes an improved algorithm C-SU (Correlation-Symmetric Uncertainty) based on Correlation Feature Selection and Symmetric Uncertainty. Due to its high stability, Random Forest is chosen as the classifier to establish the framework for the defect prediction model. To verify the validity of introduced method, it was in comparison to basic methods on datasets from eight open-source projects. The proposed feature selection method showcases superiority over other algorithms, as evidenced by the F1 score during model evaluation.