A Comparative Study of Hybrid Fault-Prone Module Prediction Models Using Association Rule and Random Forest.

Many fault-prone module prediction methods are implemented using machine learning algorithms, and the random forest is well known as the simple and powerful one. However, since the random forest uses an ensemble of decision trees, it is hard to explain why the module is predicted as “fault-prone.” In order to compensate for such a weakness, there have been studies of hybrid prediction methods combining the association rule mining technique with the random forest. In the hybrid method, a module’s fault-proneness is first assessed by the association rules. Then, when the module’s feature does not match any rules, its fault-proneness is evaluated by the random forest model. This paper focuses on how to combine the two techniques and conducts a comparative study to explore a better hybrid prediction method. The empirical results show: (1) it is better to use both association rules of “faulty” and “non-faulty” rather than using only “faulty” rules; (2) it is better to train the random forest classifiers using all data regardless of whether or not they matched association rules.