Software Fault Prediction Based on Fault Probability and Impact

Abstract

Nowadays, software tests prioritization is a crucial task. Indeed, testing exhaustively the whole software system can be very difficult, heavily time and resources consuming. Using machine learning algorithms to predict which parts of a software system are fault-prone can help testers to focus on high-risk parts of the code and improve resources allocation. This paper aims to investigate the potential of a risk-based model to predict fault-prone classes. The risk of classes is evaluated based on two factors: the probability that a class is fault-prone and its impact on the rest of the system. We used object-oriented metrics to capture the two risk factors. The risk of a class is modeled using the Euclidean distance. We built various variants of the risk-based model using a data-set from five versions of the ANT system. We used different machine learning algorithms (Naive Bayes, J48, Random Forest, Support Vector Machines, Multilayer Perceptron and Logistic Regression) to construct various models for fault and level of severity prediction. The objective was to distinguish between classes containing trivial and high severity faults. The considered model achieves good results for binary fault prediction. In addition, the overall multi-classification of severity levels is more than acceptable.