Dynamic Adversarial Feature Fusion for Tennis Win‐Rate Prediction: A Comparative Machine‐Learning Study

Traditional tennis win‐rate models rely on static technical statistics and, therefore, fail to capture the dynamic, adversarial nature of professional matches. To address this gap, we propose DAF‐ADA, a dynamic adversarial feature‐fusion framework that couples multisource sensor data (Hawk‐Eye trajectories, inertial‐sensor metrics, and plantar pressure) with a biomechanically constrained generative adversarial network for tactical‐mutation augmentation, an LSTM–attention temporal extractor, and dynamic weighting with psychological‐resilience quantification. On 320 clay‐court matches by 20 players (64,000 analysis windows) with a held‐out player split, DAF‐ADA achieved 57.1 ± 1.8% accuracy, 2.4 percentage points above the industry benchmark and on par with the strongest baseline (LSTM). Under five adversarial perturbation regimes, its mean accuracy (54.0%) was comparable to that of LSTM (54.5%), with a relative advantage under extreme serve‐speed perturbations (50.7% vs. 49.4%). These results provide a fully documented, reproducible benchmark for win‐rate prediction under player‐level generalization and show that biomechanically constrained adversarial augmentation and dynamic feature fusion yield competitive accuracy and near‐real‐time inference, while uniform robustness gains over strong sequence baselines are not obtained on this dataset.