Dynamic detection of software defects using supervised learning techniques

In software testing, automatic detection of faults and defects in software is both complex and important. There are different techniques utilized to predict future defects. Machine learning is one of the most significant techniques used to build such prediction models. In this paper, we conduct a systematic review of the supervised machine learning techniques (classifiers) that are used for software defect prediction and evaluate their performance on several benchmark datasets. We experiment with different parameter values for the classifiers and explore the usefulness of employing dimensionality reduction techniques, such as Principle Component Analysis (PCA), and Ensemble Learning techniques. The results show the effectiveness of the considered classifiers in detecting bugs. Additionally, using PCA did not have a noticeable impact on prediction systems performance while parameter tuning positively impact classifies' accuracy, especially with Artificial Neural Network (ANN). The best results are obtained by using Ensemble Learning methods such as Bagging (which achieves 95.1% accuracy with Mozilla dataset) and Voting (which achieves 93.79% accuracy with kc1 dataset).