Increasing the Prediction Quality of Software Defective Modules with Automatic Feature Engineering

Abstract

This paper reviews the main concepts related to software testing, its difficulties and the impossibility of a complete software test. Then, it proposes an approach to predict which module is defective, aiming to assure the usually limited software test resources will be wisely distributed to maximize the coverage of the modules most prone to defects. The used approach employs the recently proposed Kaizen Programming (KP) to automatically discover high-quality nonlinear combinations of the original features of a database to be used by the classification technique, replacing a human in the feature engineering process. Using a NASA open dataset with Software metrics of over 9500 modules, the experimental analysis shows that the new features can significantly boost the detection of detective modules, allowing testers to find 216% more defects than with a random module selection; this is also an improvement of 1% when compared to the original features.