Performance of Maintainability Index prediction models: a feature selection based study

Abstract

Numerous design metrics have been studied in the literature to assess software performance and future maintainability issues. Maintainability Index (MI) is an indicator of software maintenance efforts, but that can be computed only after the complete code is developed. MI is complex to calculate and hence has been estimated by several researchers using design metrics. Various prediction models are built using different sets of design metrics as independent variables. Further, researchers have used different prediction models for predicting software maintainability. To identify most prominent design metrics and the prediction models that show consistent performance irrespective of the choice of design metrics, the present empirical study is performed on datasets of 26 open source Java projects with more than one thousand and six hundred class hierarchies. Seven feature selection methods are applied to find out the most prominent metrics that are selected by majority of methods. Then four most frequently used prediction models namely, multi linear regression, multilayer perceptron, support vector regression and M5P regression tree are analyzed for their performance with respect to initial selection of design metrics. The study concludes that MIF, MaxDIT, LCC, TCC are the most significant predictors for Maintainability Index of class hierarchies. Among the all the prediction models, support vector regression model exhibits the best performance both with respect to the choice of metrics and irrespective of initial metric selection.