Basketball is a fast-paced and strategic sport where each possession involves a series of complex decisions and actions. Analyzing and understanding these intricacies is essential for effective performance evaluation. This study introduces a novel framework, Expected Immediate Value, for evaluating basketball possessions from tracking data using Graph Neural Networks. We take inspiration from the Expected Possession Value framework for predicting points scored as an immediate consequence of the current game state. We develop four specific models to enhance the interpretability and composition of metrics: xFG (probability of a shot being successful), xNAT (next action - pass or shot), xR (probability of a player receiving the ball), and xTO (likelihood of a turnover). Our approach provides a comprehensive evaluation of player movements and decisions. This framework offers deeper insights into possession dynamics and supports strategy optimization in basketball.
© Copyright 2025 Machine Learning and Data Mining for Sports Analytics. MLSA 2024. Communications in Computer and Information Science. Veröffentlicht von Springer. Alle Rechte vorbehalten.