Interpretable machine learning with SHAP for esports performance analysis of professional counter-strike players from 2012 to 2025

Performance analysis in traditional sports has become a mature area of sports science, whereas its application in esports remains relatively limited. This study applies an interpretable machine learning framework combining SHAP analysis and five classic algorithms to evaluate the performance of professional Counter-Strike players. Among the models tested via five-fold cross-validation, XGBoost demonstrated the best performance. The dataset comprises 1468 professional players from the Counter-Strike: Global Offensive period (2012–2023) and 880 from the Counter-Strike 2 period (2023–2025), with each player evaluated using 39 advanced in-game behavioral metrics categorized into seven skill dimensions (e.g., firepower, entrying, trading, clutching, opening, sniping, utility). Results show that key performance indicators such as Kills Per Round (KPR), Opening Success (OS), Rounds With a Kill (RWK), Rounds With a Multi-Kill (RWMK), Support Rounds (SR), Pistol Round Rating (PRR), and Saves Per Round Loss (SPRL) consistently ranked among the top ten features in both periods. The top six influential features of the model reveal interactions between features and exhibit a positive correlation with SHAP values. The case diagnostics further enabled precise evaluation of individual strengths and weaknesses. This exploratory study provides empirical insights and an interpretable machine learning analytical paradigm for performance evaluation and talent identification in esports.