Applying Machine Learning to Predict Software Fault Proneness Using Change Metrics, Static Code Metrics, and a Combination of Them

Predicting software fault proneness is very important as the process of fixing these faults after the release is very costly and time-consuming. In order to predict software fault proneness, many machine learning algorithms (e.g., Logistic regression, Naive Bayes, and J48) were used on several datasets, using different metrics as features. The question is what algorithm is the best under which circumstance and what metrics should be applied. Related works suggested that using change metrics leads to the highest accuracy in prediction. In addition, some algorithms perform better than others in certain circumstances. In this work, we use three machine learning algorithms (i.e., logistic regression, Naive Bayes, and J48) on three Eclipse releases (i.e., 2.0, 2.1, 3.0). The results showed that accuracy is slightly better and false positive rates are lower, when we use the reduced set of metrics compared to all change metrics set. However, the recall and the G score are better when we use the complete set of change metrics. Furthermore, J48 outperformed the other classifiers with respect to the G score for the reduced set of change metrics, as well as in almost all cases when the complete set of change metrics, static code metrics, and the combination of both were used.