Ensembled Machine Learning Approaches for Software Defect Prediction with Autoencoder Feature Selection

Software Defect Prediction models make predictions about the probability of defective modules in a software package by using historical software data like code complexity metrics and prior defect records. This allows developers to only concentrate their testing resources on discovering likely trouble spots within the code base, thus further enhancing the overall reliability of the software and reducing maintenance costs. This paper discusses different machine learning techniques such as KNN, Decision Trees, Random Forest, CatBoost, and XGBoost applied in SDP for classifying the modules of software to be defect-prone or defect-free. The paper will also discuss feature selection and outlier detection techniques including autoencoders to enhance the accuracy of such predictions. Using the PROMISE dataset as an example, these models shall hence be evaluated and showcased in order to illustrate the potential of SDP in optimizing the lifecycle of software development by possibly revamping defect detection and using resources in an economical way.