Pass2vec: Analyzing soccer players’ passing style using deep learning

Abstract

The aim of this research was to analyze the player’s pass style with enhanced accuracy using the deep learning technique. We proposed Pass2vec, a passing style descriptor that can characterize each player’s passing style by combining detailed information on passes. Pass data was extracted from the ball event data from five European football leagues in the 2017–2018 season, which was divided into training and test set. The information on location, length, and direction of passes was combined using Convolutional Autoencoder. As a result, pass vectors were generated for each player. We verified the method with the player retrieval task, which successfully retrieved 76.5% of all players in the top-20 with the descriptor and the result outperformed previous methods. Also, player similarity analysis confirmed the resemblance of players passes on three representative cases, showing the actual application and practical use of the method. The results prove that this novel method for characterizing player’s styles with improved accuracy will enable us to understand passing better for player training and recruitment.