Machine-Learning-Based Gate Selection for Accelerating Airport Gate Reassignment

Airport gate reassignment under disruptions is a large-scale, time-sensitive combinatorial optimization problem. Traditional optimization methods can deliver high-quality solutions but are often too slow for full-scale online reassignment. To address this, we propose a machine-learning-based gate-selection framework that first identifies a compact subset of gates likely to participate in a disrupted reassignment and then solves the resulting reduced problem with an optimization solver. Gate-level training records are derived from optimization-generated reassignment solutions under simulated irregularity scenarios, and the labels are gate-selection probabilities estimated from repeated solver runs. Using structured gate-level features, we formulate the prediction task as a regression problem and employ eXtreme Gradient Boosting (XGBoost). Computational experiments on data from an airport in Jiangsu Province, China demonstrate that the framework attains high recall in offline evaluation, substantially reduces online runtime, and outperforms random and heuristic greedy screening while preserving reassignment quality. The method is also robust to distributional shifts in irregularity type, irregularity ratio, occurrence time, and observation horizon. Overall, the results indicate that ML-based gate selection can substantially improve the computational tractability of airport gate reassignment while retaining the advantages of optimization-based decision making.