Decoding team performance in a self-organizing collaboration network using community structure

Abstract

When assembling a team, it is imperative to assess the ability of the team to perform the task in question and to compare the performance of potential teams. In this thesis, I investigate the predictive power of different community detection methods in determining team performance in the self-organizing Kaggle platform and find that my methodology can achieve an average accuracy of 57% when predicting the result of a competition while using no performance information to identify communities. First, I motivate our interest in team performance and why a network setting is useful, as well as present the Kaggle platform as a collaboration network of users on teams participating in competitions. Next, in order to identify communities, I applied a selection of techniques to project the Kaggle network onto a team network and applied both spectral methods and DBSCAN to identify communities of teams while remaining ignorant of their performances. Finally, I generated cross-cluster performance distributions, evaluated the significance of communities found, and calculated a predictor statistic. Using holdout validation, I test and compare the merits of the different community detection methods and find that the Cosine Similarity in conjunction with spectral methods yields the best performance and provides an average accuracy of 57% when predicting the pairwise results of a competition.