In this thesis, we work on finding the style of play in basketball using spatial data. We focus on the classification and grouping of teams based on the movement of the ball in their attack. Original data is transformed into vectors and images and used in the classification and for producing a better representation of the attack. We used random forest and neural networks for the classification and autoencoders for finding the latent data space. With the developed methods, we achieve the classification accuracy of 7.8% and get a representation with which we can describe the style of play. This representation is useful in the search for coaches and players to improve the team, and can also be used as an additional attribute for prediction of the winner.
|
