Cross-Project Defect Prediction using Supervised and Unsupervised Learning: a Replication Study

Successful software projects are now an important challenge, the main focus of the engineering community being to predict software failures based on the history of buggy classes. However, software defect prediction techniques are effective as long as there is enough data to train the prediction model. To mitigate this problem, cross-project defect prediction is used. The purpose of this investigation is two-fold. First, replicate the experiments in the original articles and second, investigate other settings regarding defect prediction with the aim of bringing new insights and results on the best approach. In this study, three supervised (Random Forest -RF, Logistic Model Tree -LMT, Naive Bayes -NB) and three unsupervised (Expectation Maximization -EM, DBSCAN, KMeans) approaches are investigated. The experiments used use preprocessed methods (normalization and feature selection). Two sets of experiments are performed considering all available features and using preselected features by Principal Component Analysis, each set of experiments being employed on both supervised and unsupervised methods. The results of the replicated experiments confirm the original findings: when using supervised methods considering all features the best method is NB, followed by RF and LMT and similar to better results when considering fewer features (with PCA); when using unsupervised methods the results are not better than the original; however, when considering fewer features the results obtained with the newly considered methods (EM, DBSCAN, KMeans) are better than in the original paper.