PUB - Publikationen an der Universität Bielefeld

To improve one's skill in two-versus-two sports, one should play regularly with varying opponents and teammates of similar skill level. However, even in small classes, it is almost impossible for a human instructor to maintain an accurate overview of each student's skill development to optimize teams and pairings, accordingly. Therefore, we propose an educational data mining approach to automated matchmaking. In particular, we trace all players' skill via the glicko2 algorithm and use the resulting skill ratings to optimize matchmaking in an integer linear programming approach. We explain the resulting matches in terms of ratings and counterfactual explanations to enhance the transparency of the system for instructors and players. In addition to the algorithm, we provide an evaluation on synthetic data and in a field study ($N=38$) conducted in a course for the fast-paced two-versus-two sport Roundnet. Our analyses showed that our proposed approach outperforms all baselines in terms of objective matchmaking performance. The subsequent field study corroborated the positive algorithmic evaluation, comparing experience from participants subjected to our proposed matchmaking approach or a random baseline. Participants' responses indicate that our approach was perceived as more trustworthy, and the explanations associated with it were deemed to be more actionable, useful, and of higher quality.