A Model Development For Software Defect Prediction: Using Feature Reduction Method

Abstract

Software is a significant element in many of the devices and systems that pervade our today‟s societies. Many software companies are building different size of software systems for various purposes to deliver the software within short time and better quality to their customers. Due to the increasing size of software data and the constraints under which the software is developed, it is too difficult to produce quality software in shorter time. Software defect is errors that are introduced by software developers (testers) and stakeholders through testing. Defect can be error, fault, bug or failure. Therefore, defect prediction before delivering the software product can contribute significantly to the success of a project in terms of; quality and cost. Software companies have various measurement methods to test their software quality. The goal of this study is to apply different software data set before preprocessing and feature reduction in machine learning algorithms or methods for predicting the status of software defect from the NASA MDP data set. Especially, the researcher has gave emphasize on a model development for software defect prediction using feature reduction method plus prototype system for identification of faults. There were some methodologies, including initial literature review up to evaluation of the performance and developing the proposed model. The researcher gave more attention for both before preprocessing and after feature reduction of the machine learning approach by using four data set,CM1,KC1,KC2,PC1. Machine learning algorithms such as NBTree, Logistic and NB are used for the experimentation through Weka machine learning tool to build the prediction model. The model accuracy is tested using the 10-fold cross validation mode and prest software tool has been used to show the extraction of software features. The best performing methods is identified by comparing their accuracy, specificity and execution time of build the models. Based on performance evaluator the best algorithm was found to be the(NBT) classifier, in PC1 data set feature reduction the researcher was gets 93.5% accuracy from the experiment. As a result the researchers identified that the feature selection should be tested before different software data set testing or maintaining taken place by testers or developers. Keywords:- Software Defect Prediction, Feature Selection, Software Metrics, Machine Learning, Model, Data set, Preprocessing, Performance, Accuracy.