Bayesian Prediction of Fault-Proneness of Agile-Developed Object-Oriented System

Abstract

Logistic regression (LR) and naïve Bayes (NB) extensively used for prediction of fault-proneness assume linear addition and independence that often cannot hold in practice. Hence, we propose a Bayesian network (BN) model with incorporation of data mining techniques as an integrative approach. Compared with LR and NB, BN provides a flexible modeling framework, thus avoiding the corresponding assumptions. Using the static metrics such as Chidamber and Kemerer’s (C-K) suite and complexity as predictors, the differences in performance between LR, NB and BN models were examined for fault-proneness prediction at the class level in continual releases (five versions) of Rhino, an open-source implementation of JavaScript, developed using the agile process. By cross validation and independent test of continual versions, we conclude that the proposed BN can achieve a better prediction than LR and NB for the agile software due to its flexible modeling framework and incorporation of multiple sophisticated learning algorithms.