A Novel Machine Learning Method for Fault Prediction and Reliability in Software Systems

Predicting the emergence of a software flaw during its early phases of development requires software fault prediction. Early predictions of faults can help software achieve both greater reliability and lower costs. Prediction strategies that improve software quality have recently been developed using a variety of machine learning methodologies. This paper presents a novel machine learning methodology for software failure prediction utilising PROMISE's CM1 data. The method makes use of ADASYN to handle imbalanced classes, uses SelectKBest for feature selection and applies Min-Max scaling for normalization. A Deep Neural Network (DNN) classifier is trained and tested by using ReLU activation and SoftMax output. Our analysis showed that the model reaches high accuracy, precision, recall, F1score and ROC scores, indicating good fault prediction. The outcome of experiments suggests the model is good at minimizing false positives and detecting faults, scoring an accuracy of 97.67%, 98% in precision and 96% in recall. An F-score of 97% and Area Under the Curve (AUC) of the same value highlight that the model shows good and reliable results when classifying binary tasks. It appears that deep learning can greatly aid in defect prediction for software engineering, resulting in better quality assurance, less maintenance and other good outcomes.