SOFTWARE DEFECT PREDICTION USING HOMOGENEOUS AND HETEROGENEOUS ENSEMBLE TECHNIQUES

Abstract

Many researchers have already been working in the field of defect prediction in software using some machine learning algorithms. Their results vary from dataset to dataset. These algorithms give inconsistent output for predicting defects in a random software project. Researchers have not decided which machine learning algorithm is best suitable for correctly predicting the defects in software so recent developments in machine learning introduce ensembling methods to predict defects. Ensembling takes the advantages of different techniques to give a better prediction of defects compared to individual base models.

The major objective for work is building the ensemble of various classification methods to predict the defects in the given software module and compare the results of the ensemble with an individual classification technique. We have used naive bayes classifiers, logistic regression, k- nearest neighbors, support vector machine, decision trees for implementation and then choose the best three classification techniques to build the ensemble, and data sets are collected from publicly available repositories. Here we have used heterogeneous ensemble techniques such as voting, stacking and homogeneous ensemble techniques such as bagging and boosting for prediction. Also heterogeneous version of bagging and boosting is used. All the six techniques are implemented and compared using the various performance metrics. Area under ROC curve (AUC) is used to analyze the prediction performance and to check the statistical significance of the results of different models, Friedman test is used. The results show that the ensemble method improves the prediction performance as compared to individual classifiers.