Comparison of threshold identification techniques for object-oriented software metrics

Abstract

Quality assurance is a continuous process throughout the project lifecycle from inception till post-delivery. Software metrics are tools to help developers in achieving software quality objectives. Software metrics are used to predict the fault-proneness of classes in software using machine-learning and statistical techniques. However, these methodologies are difficult for daily tasks. Simpler and on the fly methodologies such as threshold values are needed. Metric thresholds can be used to control software quality and to recommend improvements on software code. Thresholds detect the parts of software that need more verification and validation. Many threshold identification techniques were proposed in previous research. However, the techniques do not provide consistent thresholds. The authors compare eight threshold identification techniques to diagnose software fault-proneness. The eight techniques are derived from diagnosis measures such as specificity, sensitivity, recall and precision. Five threshold identification techniques have derived thresholds that are skewed and have large standard deviations. Only three techniques are selected for threshold identification based on consistency and variation in selecting thresholds of software metrics in the systems under study. These techniques find thresholds that have the least variation among the studied techniques. The median of the 11 systems is selected as a representative of all thresholds.