Comparing machine learning techniques for software requirements risk prediction

Abstract

Software requirements are the most critical phase focused on documenting, eliciting, and maintaining the stakeholders’ requirements. Risk identification and analysis are preemptive actions designed to anticipate and prepare for potential issues. Usually, this classification of risks is done manually, a practice that the personal judgment of the risk analyst or the project manager might influence. Machine learning (ML) techniques were proposed to predict the risk level in software requirements. The techniques used were logistic regression (LR), multilayer perceptron (MLP) neural network, support vector machine (SVM), decision tree (DT), naive bayes, and random forest (RF). Each model was trained and tested using cross-validation with k-folds, each with its respective parameters, to provide optimal results. Finally, they were compared based on precision, accuracy, and recall metrics. Statistical tests were performed to determine if there were significant differences between the different ML techniques used to classify risks. The results concluded that the DT and RF are the techniques that best predict the risk level in software requirements.