A NOVEL APPROACH TO IMPROVE SOFTWARE DEFECT PREDICTION ACCURACY USING MACHINE LEARNING

Abstract

Defect prediction is a prominent field within the software engineering community. In order to
ensure the program's success, it is crucial to minimize the disparity between software engineering and data
mining. Software defect prediction anticipates the occurrence of source code problems prior to the testing
process. Various techniques, including clustering, statistical approaches, mixed algorithms, neural networkbased
metrics, black box testing, white box testing, and machine learning, are often used to forecast software
problems and analyze their impact. This study introduces the novel use of feature selection to enhance the
accuracy of machine learning classifiers in predicting faults. The aim of this project is to enhance the precision
of defect prediction in five NASA datasets, namely CM1, JM1, KC2, KC1, and PC1. The NASA data sets are
publicly accessible. This research employs feature selection technique in conjunction with various machinelearning
techniques, namely Random Forest, Logistic Regression, Multilayer Perceptron, Bayesian Net, Rule
ZeroR, J48, Lazy IBK, Support Vector Machine, Neural Networks, and Decision Stump. The objective is to
enhance defect prediction accuracy significantly when compared to the scenario where feature selection is not
applied (WOFS). The research workbench utilizes a machine-learning program known as WEKA (Waikato
Environment for Knowledge Analysis) to enhance and preprocess data, as well as implement the specified
classifiers. A tiny tab statistics tool is used for evaluating statistical studies. The research findings indicate that
the accuracy of defect prediction is enhanced while using feature selection (WFS) compared to the accuracy of
WOFS.