Software quality assurance is an area that deals with software defect prediction also. Identifying and eliminating defects is a crucial task that helps organizations deliver quality software products to customers. Machine learning approaches help in identifying software modules that are defective and which are not defective. The existing software defect prediction datasets contain data with features that could classify projects are defective or not. The machine learning model’s performance will be degraded with the existence of noisy attributes and class imbalance problems. In this work, we propose a ROS-KPCA-SG model (Random Over Sampling-Kernel Principal Component Analysis-Staked Generalization Model) model to solve the noisy dataset and class imbalance problems and to improve the software defect prediction accuracy. The performance of the ROS-KPCA-SG model is compared with individual models with different combinations of sampling techniques. The results show the proposed ROS-KPCA-SG model solves the problems and gives better performance than other models. The AUC-ROC score is between 0.9 and 1 for the ROS-KPCA-SG model on all the datasets, and the accuracy is near to 90% and above which is a higher value than other models. The proposed model gives accuracy on datasets CM1 is 98%; JM1 is 89%; PC1 is 98%; KC1 is 92% and with KC2 dataset is 89%.