Using the Silhouette Coefficient for Representative Search of Team Tactics in Noisy Data

Automatically recognizing team tactics based on spatiotemporal data is challenging. Deep Learning approaches have been proposed in this area but require a tremendous amount of manual work to create training and test data. This paper presents a clustering approach to reduce the needed manual effort significantly. A method is described to transform the spatiotemporal data into a canonical form that allows to efficiently apply clustering techniques. Since noise cannot be avoided in the given application context, the silhouette coefficient is applied to filter clusters considered to be noisy in a cluster technique independent way. Then, a variant of the silhouette coefficient is introduced as an indicator regarding the overall cluster model quality which allows to select the optimal clustering technique as well as the optimal set of cluster technique parameters for the given application context.