Detection of Fault-Prone Classes Using Logistic Regression Based Object-Oriented Metrics Thresholds

Background: In the plethora of studies, the objectorientedmetrics have been empirically validated to assess thedesign properties and quantify the high-level quality attributessuch as fault-proneness, either at the method or class granularitylevels of software. Motivation: A more precise value of an objectorientedmetric can be used as an indicator for the developers tomake the informed decisions regarding the detection of designflaws and classify the fault-proneness classes. Method: Benderused an approach in the domain of epidemiology studies to derivethe threshold values for the risk factors. In our study, we followthe Bender's approach and propose a model to derive thethresholds for a set of software design metrics via non-linearfunctions, which are described through logistic regressioncoefficients. Subsequently, we perform four types of analysis andthree experiments in order to evaluate and compare theeffectiveness of derived thresholds in the domain of classificationof fault proneness classes. We use the Precision, Recall, Fmeasureand classification accuracy performance measures toassess the effectiveness of derived metrics thresholds. Results: Wecompare the derive threshold values of DIT, CA, LCOM andNPM metrics with their existing data distribution basedthreshold values, and observed the significant increase in theclassification accuracy of fault-prone classes. For example, DIT(27%), Ca (2%), NPM (2%) and LCOM (15%) for the Ant-1.5project. Conclusion: The analysis results suggest that theproposed model can be applied to derive the thresholds of otherobject-oriented metrics which present either with or withoutheavy-tailed distribution, however, the proposed model to derivethresholds cannot generalize for all the systems due to variationin data characteristics.