Development of Defence and Offence Play Items for Deep Learning Model of Offence Play Analysis in Soccer Game
Hiroki Matsuoka, Yasuhiro Tahara, Kozue Ando and Takahiko Nishijima
[Received September 9, 2019; Accepted June 22, 2020]
The purpose of this study was to develop offence and defence tactical play items for a deep learning model of tactical play analysis. The procedures used in this study included the following three steps. First, features of offence and defence tactical play items were developed. Then, long short-term memory (LSTM) was architected. Finally, tactical plays were predicted by the model. The ball touch data and the tracking data from two official soccer games in the J-League 2016 season were used. The ball touch data, recorded player actions in text such as passes and shots with time-series order, and the tracking data of players, were used to construct thirty-one tactical play items. For the deep learning model, LSTM was used. LSTM allows the analysis of time-series text data. 6,444 sequential plays were used. The highly accurate tactical play predicted from LSTM was the feed after tactical play which started at low press defence and finished at GK ball catch (8 correct predictions of 8 frequencies in the data). In conclusion, all 31 items measuring offence and defence tactical play in soccer games constructed from the ball touch data and the tracking data are the feature items used to analyse tactical plays using LSTM.
Keywords: machine learning, long-short term memory, ball touch data, tracking data
[Football Science Vol.17, 69-85, 2020]