We show that pretraining transformer models improves the performance on supervised classification of tracking data from elite soccer. Specifically, we propose a novel self-supervised masked autoencoder for multiagent trajectories. In contrast to related work, our approach is significantly simpler, has no necessity for handcrafted features and inherently allows for permutation invariance in downstream tasks.
© Copyright 2024 Machine Learning and Data Mining for Sports Analytics. MLSA 2023. Communications in Computer and Information Science. Published by Springer. All rights reserved.