Clarity From Similarity: A Data-Driven Approach to Football Player Role Classification

Abstract

Football players have traditionally been categorized by broad positions by their positions on the pitch, which often fail to reflect the nuanced differences in individual playing styles. This thesis addresses the shortcomings of position-based labels by introducing a data-driven framework that classifies players into specific role categories based on detailed performance metrics. The methodology leverages domain knowledge and previous research to predefine a set of prototypical role archetypes for each positional group and classifies players by measuring their statistical similarity to these archetypes. This similarity-based approach results in continuous role profiles for players - offering a more nuanced, interpretable representation of on-field roles than traditional position labels do. The evaluation involved comparing model classifications against expert assessments to determine the model’s reliability. Results showed that, out of 52 instances tested, the model outperformed a random assignment baseline 41 times, matched it on 4 occasions, and underperformed in 7 instances; these latter cases were highlighted as areas for improvement in future research. Finally, we present a potential integration of the model with FotMob’s platform.