Automatic Classification of Scientific Groups as Productive: An Approach based on Motif Analysis

One of the key aspects instrumental in the advancement of science relates to "team science," or in other words "group" collaborations. There have been extensive studies analyzing various statistical properties of collaborations of individual or pairs of authors. However, the number of studies pertaining to groups/teams of scientists working together is limited in number. In this paper, we set an objective to study the productivity of group collaborations where groups are represented as small substructures usually termed as network motifs in the literature. A preliminary observation is that star-like motifs have the largest productivity (defined as a function of citation count) followed by 4-cliques. We then introduce a bunch of features and study their individual relations with the productivity of a team. Building on these observations, we develop a supervised classification model that can automatically distinguish the highly productive teams from the low productive ones based on the set of identified features. The accuracy of the classification is 82% on an average for all the motifs with the accuracy reaching as high as 95% for 4-cliques. Finally, we present a detailed analysis of the time-transition behavior of different motifs along with some of the real world highly productive motifs found in our dataset. This empirical study is a first step toward the development of a full-fledged recommendation system that can predict how productive a team would be in the future.