Discovering Tactical Weaknesses in Football Using Unsupervised Machine Learning on Event Data

Abstract

The aim of this thesis is to explore how unsupervised machine learning can detect tactical weaknesses in football, specifically in defensive structures and defensive transitions following possession loss. By extracting spatial features like team centroid, width, depth, compactness and distance to the ball from Metrica Sports open data, team formation was represented. Principal Component Analysis (PCA) was used for data visualization, and K-Means and HDBSCAN were used to identify tactical phases and team formations. Isolation Forest was applied for anomaly detection to detect outliers that might represent potential tactical weaknesses such as poor pressing, stretched formations, and slow defensive transitions. The findings reveal that tactical weaknesses tend to occur in temporary moments rather than over the whole game. This research shows that unsupervised machine learning can be used to generate tactical insights for coaches and analysts without the need for labelled data.