Improving performance of software requirements classification with ensemble learning method

In recent years, the widespread use of developing technology and environmental factors such as pandemic, which negatively affects society, make software applications an indispensable part of our lives. This situation causes a great increase in the number of software projects that meet with users. In parallel with the increase in the number of software projects, it becomes more important for companies to put forward efficient and successful software projects at the point of competition. Success in software projects is evaluated in terms of the ability to fully fulfill the features and functions requested by the customer within the determined budget and time. One of the most important steps to be considered for the successful conclusion of software projects is software requirement analysis studies. In the requirements analysis step, which is the first step in the project life cycle, the more accurately the requirements are determined, and the more accurately these determined requirements are analyzed, the next steps of the project are carried out in a more controlled manner. In software projects, especially functional and non-functional requirements need to be handled carefully. In this study, the performances of various machine learning algorithms are tested in order to reduce the time spent and make people independent by using data science in the classification of functional and non-functional requirements. Using the PROMISE-NFR dataset, the individual results of the machine learning algorithms were evaluated with the tests performed. The performances of the algorithms are compared using the accuracy metric. As a result of the comparison, the highest performance value was achieved with the Naive Bayes algorithm, with an accuracy value of 0.8992. Ensemble methods were used with the aim of increasing the singular performance of the algorithms and the double and triple interactions of the algorithms with the Vote application were examined. For double and triple evaluations, ten machine learning algorithms in each group were determined and the relationships between them were examined. When the performances of the algorithms considered as binary were evaluated, the desired success could not be achieved. When the performances of the algorithms, which are considered triples, are evaluated, the best result with an accuracy value of 0.9024 was obtained with the use of SMO, PART, and Random Forest algorithms together. As a result, it has been determined that ensemble learning methods give better results than single machine learning methods.