EMPIRICAL VALIDATION OF PROCESS METRICS TO CHECK THE PREDICTIVE PERFORMANCE OF CLASSIFICATION AND ENSEMBLE METHODS

Abstract

Software defects have always been considered a major problem in the software industry and for software engineers, early detection improves software performance and reduces faults, time, and cost. In order to predict defects in software, many researchers have been used classification and ensemble techniques. Different dataset produces different results. In this research, we have evaluated the prediction accuracy of classification and ensemble approaches using 3 distinct models: combined model of static code and process metrics, model containing process metrics, and model containing static code metrics. In simple terms, we can say that these 3 models have different independent variables and dependent variables are the actual values of bugs which is the same. We have used NB, LR, KNN, SVM, DT as classification approaches and stacking, voting, bagging, and boosting as ensemble approach for implementation. The dataset was gathered from the publicly available repository. AUC metric was used to examine the prediction performance of classification and ensemble techniques. Additionally, the statistical significance of the results obtained from various models was assessed using the Friedman and Nemenyi post hoc test. The result of this study demonstrates that the use of process metrics in predicting the defects in software produces effective outcomes.