Classification and prediction of software cost through fuzzy decision trees

Abstract

This work addresses the issue of software effort prediction via fuzzy decision trees generated using historical project data samples. Moreover, the effect that various numerical and nominal project characteristics used as predictors have on software development effort is investigated utilizing the classification rules extracted. The approach attempts to classify successfully past project data into homogeneous clusters to provide accurate and reliable cost estimates within each cluster. CHAID and CART algorithms are applied on approximately 1000 project cost data records which were analyzed, pre-processed and used for generating fuzzy decision tree instances, followed by an evaluation method assessing prediction accuracy achieved by the classification rules produced. Even though the experimentation follows a heuristic approach, the trees built were found to fit the data properly, while the predicted effort values approximate well the actual effort. © 2009 Springer Berlin Heidelberg.