Related Experiment Videos
Optimizing football analytics: Dimensionality reduction meets machine learning for offensive and defensive player
Rodrigo Yáñez-Sepúlveda1,2, Aldo Vásquez-Bonilla3, Rodrigo Olivares4
1Faculty of Education and Humanities, Universidad Andres Bello, Viña del Mar, Chile.
Background:
This study focuses on the quantitative analysis of football players' performance based on a set of performance data.
Objective:
The objective of this study was to develop a multivariate analytical model to identify, quantify, and predict the most relevant factors that determine the offensive and defensive profiles of professional football players, thereby optimizing functional segmentation, performance analysis, and decision-making processes in the contexts of scouting and data-based tactical planning.
Materials And Methods:
A dataset was compiled through automated extraction (web scraping) via the Python programming language and its specialized libraries. The full dataset comprised 2,689 player-league records corresponding to 2,530 unique players from Europe's five major UEFA-affiliated national leagues during the 2022-2023 season (159 players who transferred between two leagues mid-season contributed one record per league, 2,530 + 159 = 2,689). After applying a minimum-eligibility filter of >450 minutes played to ensure stable per-90 metrics, 1,624 player-league records (1,620 unique players) were retained for the modeling workflow. Dimensionality reduction techniques (PCA) and explainable machine learning (Shapley Additive exPlanations [SHAP]) with predictive modeling were applied. Eight algorithms were trained and compared to predict PC1 Offensive and PC1 Defensive.
Results:
The First Offensive Principal Component (PC1 Offensive) synthesizes actions that generate scoring opportunities, such as actions prior to shots (SCA_p90, 13.3%), progressive receptions (RecProg_p90, 13.0%) and total shots (Shots_p90, 12.7%). The first-defensive principal component (PC1 Defensive) highlights tackles (Tkl_p90, 19.1%), blocks (15.6%) and interceptions (15.6%).
Conclusions:
Finally, the inclusion of SHAP allows for an in-depth interpretation of the models, identifying the key statistical factors that influence players' offensive and defensive contributions, which is valuable for scouting, tactical analysis and player development.