Using neural network interpretability to understand outcomes in women’s 3 x 3 basketball

Abstract

Despite the inclusion of 3x3 basketball in the Olympic Games, research on this topic remains sparse, especially concerning women's 3x3 basketball. This study aimed to understand game outcomes in the FIBA 3x3 women's professional circuit. Data was sourced from the official FIBA 3x3 website, encompassing 15 indicators and 2 contextual variables from 987 matches across four seasons. All games were classified into balanced and unbalanced categories. The Multilayer Perceptron neural network outperformed discriminant analysis in both balanced and unbalanced games, achieving classification accuracy exceeding 85%. To interpret the neural network's predictions, we calculated SHAP values, revealing that one-point field goal made and defensive rebounds were the key performance indicators. In balanced games, free-throw made and ball possession contributed significantly to the classification of winning and losing teams, while team fouls and turnovers were instrumental in distinguishing outcomes in unbalanced games. This study provides valuable insights into game outcomes in women's 3x3 basketball.