An Adaptive Approach for defect Prediction using Dynamic Classifiers Selection

Abstract

Software Defect Prediction (SDP) approaches use learning methods to classify classes/module/files into the defective or non-defective or provide the possibility that a class can show faulty behaviors in the future. Since there are several classifiers that can give optimal results using ensemble learning methods, they are developed to estimate the defect-proneness of a class by combining prediction outcomes obtained from different classifiers. We are employing an ensemble learning technique and building an adaptive approach for performing bug prediction by dynamically selecting one classifier from a set of machine learning classifiers, that predicts if a class is bug prone or not based on characteristics of static software metrics captured for class. We are proposing a new Adaptive Approach for Bug Prediction using Dynamically Classifier Selection (ADCS) which dynamically selects the best base learning that better predicts bug-proneness of a class based on characteristics of the class. We have used datasets obtained from the PROMISE repository (developed by NASA) for 30 Software system. Our results indicates that ADCS perform better compared to 5 different classifiers which are used to predict bug-proneness independently and when Validation and Voting (VV) ensemble technique used to combine classifiers output with majority voting. We found that the ADCS outperforms in 26 projects and avoids class imbalance and overfitting problems.