Evaluating the Impact of Sampling-Based Nonlinear Manifold Detection Model on Software Defect Prediction Problem

Accurate prediction of defects is considered an essential factor, depending mainly on how efficiently testing of different prediction models has been done. Earlier, most of the models were restricted to the use of feature selection methods that had limited effects in solving this problem in initial stage of software development. To overcome it, the application of software defect prediction model using modern nonlinear manifold detection (nonlinear MD) combined with SMOTE using four machine learning classification approaches has been proposed in a way that the challenging task of defect prediction has been categorized as problem of high-dimensional datasets, problem of imbalanced class, and identification of most relevant and effective software attributes. Then, statistically evaluated and compared performance of prediction model with or without SMOTE-nonlinear MD approaches and results validated that proposed SMOTE-nonlinear MD approach prediction model predicts defect with better accuracy than others using RMSE, accuracy, and area under the curve.