Capturing Intransitive Dominance in Tennis Forecasting: A Graph Neural Network Approach
Researchers have developed a novel graph neural network (GNN) approach to forecast tennis matches by explicitly modeling intransitive player dominance. This method represents players as nodes and match outcomes as directed edges, capturing relationships where player A beats B, B beats C, and C beats A. While the GNN model achieved 65.7% accuracy and a 0.214 Brier score, it did not outperform established systems like Weighted Elo in unconditional accuracy. However, a combined forecast incorporating the GNN's complementary information significantly surpassed Weighted Elo, particularly on intransitive matchups. AI
IMPACT This research introduces a novel graph-based approach for sports forecasting, potentially improving prediction accuracy by incorporating complex player dynamics.