A multiclass classification model to estimate Agile user stories

Abstract

Effort estimation in Agile projects is extremely important in order to prevent schedule overrun, late product delivery and over-budgeting among others. These estimations are usually performed manually by making use of estimation techniques such as planning poker and expert judgment. However, there are several constraints associated with these techniques since estimation requires the presence of the whole project team and there might be divergence of opinions. The main objective of this study was to estimate Agile user stories using multiclass classification techniques. Three classifiers (Naïve Bayes, Decision Tree and Support Vector Machine) have been applied to user stories to estimate effort in terms of Story Points. Three machine learning models were generated and evaluated to determine the best algorithm by comparing metrics such as accuracy, error rate, precision, recall and f-measure. Using a single train:test sample in the ratio of 100:25, Naïve Bayes, Decision Tree and Support Vector Machine gave an accuracy of 64%, 80% and 100% respectively. Experiments performed showed that the accuracy of Support Vector Machine was better than the other two algorithms in estimating user stories.