Player2Vec and other Probabilistic Models

Framework for Evaluating Player Performance in Soccer

Abstract

Over the last decade, sports analysis has evolved into an increasingly mathematical and sophisticated perspective. Applications such as the use of spatial analysis in Basketball and Brentford’s statistical research with Smartodds are clear examples of the growing trend in this field. Baseball, long the sport of choice for analytics, has undergone a profound transformation with the implementation of Sabermetrics. The introduction of advanced analytics tools has produced positive results for many teams, highlighting the value of studying specific metrics within each sport.
This development focuses on soccer, a sport in which previous analyses have mostly concentrated on predicting match results and improving team performance as a whole. However, this work proposes a different approach by analyzing the relationship between the player as an individual and the respective player training.
Specifically, we explore the impact of players on ball possession and team shots from a probabilistic perspective. Starting from the probability of kicking before losing the ball (PSL) proposed in the developing paper “Soccer Networks”, and its demonstrated positive influence on team results, we propose a process to compare the impact of players on PSL and consequently on team performance.
We succeeded in formulating a methodology to study the PSL distribution of a team for which we propose a series of methods and metrics to compare the performance of two formations of players. We developed vector representation of the players (Embeddings), called Player2Vec, based on the player graph proposed in the same paper “Soccer Networks”. The latter allows us to develop predictive models about the performance of players in a team. Our final model is 58.99% better at predicting player performance than assuming previous distributions as priors, thus outperforming simpler Bayesian models.