Using Software Metrics Thresholds to Predict Fault-Prone Classes in Object-Oriented Software
Alexandre Boucher, Mourad Badri · 2016
Most code-based quality measurement approaches are based, at least partially, on values of multiple source code metrics. A class will often be classified as being of poor quality if the values of its metrics are above given thresholds, which are different from one metric to another. The metrics thresholds are calculated using various techniques. In this paper, we investigated two specific techniques: ROC curves and Alves rankings. These techniques are supposed to give metrics thresholds which are practical for code quality measurements or even for fault-proneness prediction. However, Alves Rankings technique has not been validated as being a good choice for fault-proneness prediction, and ROC curves only partially on few datasets. Fault-proneness prediction is an important field of software engineering, as it can be used by developers and testers as a test effort indication to prioritize tests. This will allow a better allocation of resources, reducing therefore testing time and costs, and an improvement of the effectiveness of testing by testing more intensively the components that are likely more fault-prone. In this paper, we wanted to compare empirically the selected threshold calculation methods used as part of fault-proneness prediction techniques. We also used a machine learning technique (Bayes Network) as a baseline for comparison. Thresholds have been calculated for different object-oriented metrics using four different datasets obtained from the PROMISE Repository and another one based on the Eclipse project.