Classification of Group Potency Levels of Software Development Student Teams

Abstract

This paper describes the use of an automatic classifier to model group potency levels within software development projects. A set of machine learning experiments that looked at different group characteristics and various collaboration measures extracted from a team’s communication activities were used to predict overall group potency levels. These textual communication exchanges were collected from three software development projects involving students living in the US, Turkey and Panama. Based on the group potency literature, group-level measures such as skill diversity, cohesion, and collaboration were developed and then collected for each team. A regression analysis was originally performed on the continuous group potency values to test the relationships between the group-level measures and group potency levels. This method, however, proved to be ineffective. As a result, the group potency values were converted into binary labels and the relationships between the group-level measures and group potency were re-analyzed using machine learning classifiers. Results of this new analysis indicated an improvement in the accuracy of the model. Thus, we were able to successfully characterize teams as having either low or high potency levels. Such information can prove useful to both managers and leaders of teams in any setting.