Position data in team sports, like handball, allow for novel approaches of quantitative analysis. Recent work used notational analysis of domain experts to find performance indicators related to success. The aim of this study was to scale expert knowledge using machine learning to extract counter attacks and position attacks from a data set of 539 elite-level handball matches. Videos from 10 games were analyzed. Each ball possession phase was labeled as either counter or position attack. The labels were used to train a graph-based deep neural network. The trained network was used to extract 62,648 position attacks and 19,262 counter attacks. The results show that counter attacks are shorter and players have higher mean velocities. However, distance covered are similar in counter attacks and position attacks. Additionally, winning teams attempt more then 50% of their counter attacks in the first half, while losing teams tend to have more counter attacks in the second half. The results of this study show that machine learning can be used to apply expert knowledge to large data sets and give novel insights into the characteristics of counter attacks and position attacks as well as tactical behavior of winning and losing teams.
© Copyright 2023 13th World Congress of Performance Analysis of Sport and 13th International Symposium on Computer Science in Sport. IACSS&ISPAS 2022. Veröffentlicht von Springer. Alle Rechte vorbehalten.