Football player feature learning using graphs for recommendation and retrieval

Abstract

Football analytics is a field that has grown tremendously over the years as technology for capturing data in sporting events has improved. During the transfer season, scouts and football managers must conduct numerous analyses in order to build strong and successful teams. In order to succeed and win matches, managers' in-game decisions are critical. A substitute is a player who enters the field in place of another player during a football game. This thesis proposes a recommendation system to assist football managers in making substitution decisions. The proposed system utilizes passing networks of teams to learn feature embeddings of football players. Using the extracted player embeddings, a k-nearest neighbors model, an XGBoost model and an artificial neural network model were trained to learn the relationship between suitable player pairs that can be substituted during the match. The model recommends the most suitable football player to substitute for a given input player. The main objective of employing passing graphs to gain knowledge of key player aspects is that the features extracted are unbiased. To prove the validity of this approach, two feature sets containing player attributes assigned by the FIFA19 game were used as baselines. The experiments demonstrated that using features extracted from passing graphs for player substitute recommendations yields models with higher recommendation performance when compared to the baseline models. Aside from decisions made during football games, managers are also responsible for decisions made off the field, such as selecting players to transfer to the team. Finding the most similar and suitable football players to fill a vacant position left on the team by another player who has left the club is essential for keeping the team dynamics together. A similar player retrieval system was developed for this thesis. The model employs egocentric passing networks calculated individually for each player, which take into account the passing relations of the nearest players as well as centrality values of players, which describe the players' roles in the network. It has been demonstrated that when combined with the baseline model's features namely player pass length, pass direction, and pass location, the proposed new features improve the performance of the player retrieval task. The experiments showed that using passing graphs of football teams to learn player representations can be beneficial for producing better performing recommendation and retrieval models.