Discriminative efficiency methodology for validating software quality classification models

The use of discriminant analysis for controlling software quality using complexity metrics is explored. One of the main problems in discriminant analysis involves selecting the principal components among complexity metrics that affect the classification of software modules into several groups based on quality. This paper focuses on the use of Akaike information criterion (AIC) procedures for selecting these principal components. These procedures are based on the residual sum of squares obtained from the maximum likelihood estimation and the number of complexity metrics. The validity of using AIC procedures is confirmed by Wilks A -criterion procedures, taking the discriminant efficiencies (i.e., F-test values) of models and those of complexity measures into account. Experiments show that AIC procedures can identify the model that has the highest discriminant efficiency with a low misclassification ratio and that the maximum F-test value model is an approximate solution of the minimum AIC estimated model.