Learning football player features using graph embeddings for player recommendation system

Abstract

Football analytics is a field that has been growing incredibly over the years thanks to the improvement of technologies capturing data in sports events. Outcomes of football matches are highly affected by the in-game decisions of football manager such as defending and attacking strategies or substituting particular football players. That is why football player recommendation is an important decision making task to gain the best results from a football match. To assist the football managers in this decision making process, a system that recommends the most suitable football player to replace a certain player is proposed. Our proposed model utilizes passing information during a game to learn feature embeddings of football players. Using the learned feature embeddings, a k-nearest neighbors (k-NN) model, an XGBoost model and an artificial neural network (ANN) model are trained to recommend the most similar and suitable replacement for a football player. The novelty of this recommendation system is that learned embeddings generate high-quality representations of football players which yield high performance for player recommendation when a replacement is needed.