NASA promise veri setlerinde derin ve makine öğrenme modellerinin yazılım hata tahmini performansının izlenmesi

In an era where software reliability and quality assurance have gained paramount importance, this study employs advanced machine learning models to predict software defects, thereby contributing to a refined understanding of their potential applications in enhancing software reliability. The focal point of the investigation is the PROMISE20 dataset, a collection of data from various NASA software projects. This dataset is segmented into three sub-datasets (CM1, JM1, KC1), with each instance marked by a binary dependent variable (indicating defect status) and independent variables based on Halstead and McCabe static code metrics. The study undertakes a comparative analysis between deep learning models, specifically LSTM and LSTM-GRU, and traditional machine learning models such as the XGBoost Classifier. Their proficiency in predicting software defects is gauged by assessing their accuracy and F1-scores. Upon examination, LSTM and LSTM-GRU deep learning models outperform with superior predictive performance, demonstrating accuracy rates of 88% and F1-scores of 0.89 and 0.90, respectively. In the realm of traditional machine learning, the XGBoost Classifier emerged as the top performer, boasting an accuracy rate of 0.88. However, the findings also underscore the need for further exploration. The study points to the necessity of examining additional datasets, exploring diverse models, optimizing model hyperparameters, and enhancing model interpretability to ascertain the optimal choice of model for software defect prediction. This research enriches the ongoing discourse in software reliability and defect prediction, offering a robust foundation for future investigations in the field of software defect prediction using machine learning.