[2H6-OS-8b] 移動系列のデータマイニングと機械学習

[2H6-OS-8b-03] Analysis using a generalized defensive evaluation based on soccer event prediction probability

〇Rikuhei Umemoto¹, Kazushi Tsutsui¹, Keisuke Fujii^1,2,3 (1. Nagoya University, 2. RIKEN, 3. PRESTO)

Keywords:sports, machine learning , data mining

Analysis of defenses in team sports is generally difficult due to limited event data. In soccer, a method has been proposed to evaluate team defense by using the positional relationship between all players and the ball to predict ball gains and effective attacks. However, previous studies did not consider the importance of events, assumed complete observation of all 22 players, and did not fully investigate the effects of diversity such as nationality and gender. In this study, we propose a general evaluation method for defensive teams by scaling the predicted probability of an event by goals scored and goals conceded. Using open-source data including positional data about all players in the broadcast video frames of UEFA EURO 2020 and UEFA Women's EURO 2022 soccer matches, we investigated the impact of the number of players on prediction and validated our method through match analysis. The results showed that information on all players was not necessary for predictions regarding effective attacks, goals scored, and goals conceded, but information on three to four players for each offense and defense was necessary for predictions regarding ball gains. The game analysis allowed us to explain the defensive excellence of the teams that reached the final tournament of UEFA EURO 2020.