Application of quantum machine learning in NBA game outcome analysis

With the rapid integration of sports data science and intelligent computing, outcome analysis modeling for professional competitions has become a key scientific problem. Conventional machine learning models have demonstrated strong performance on structured sports data; however, investigating more expressive feature representation strategies for complex and nonlinear learning problems remains an important research direction. This work explores the application of quantum machine learning to NBA game outcome classification analysis by comparing quantum kernel models with classical machine learning models, systematically evaluating the potential capabilities of existing QSVM methods in sports analytics, and employing SHAP-based interpretation methods to compare feature contribution patterns between quantum and classical models. In this study, we employ a domain-adapted feature encoding strategy(min-max phase mapping to \([0,π ]\) ), kernel centering with adaptive calibration, and a comprehensive comparison of multiple quantum circuit families, including 14 different Pauli combinations, three circuit depths, and two entanglement topologies. With an AUC of 0.915 and accuracy of 0.834, the optimal QSVM performs competitively against classical SVM (AUC 0.942). Robustness evaluation under depolarizing noise yields an AUC of 0.814, confirming the model’s viability for near-term quantum hardware. SHAP analysis reveals consistent feature attributions across classical and quantum models, indicating QSVMs capture meaningful basketball structure. These results establish a reproducible benchmark for quantum-enhanced sports analytics under NISQ constraints.