Generative Augmentation of Imbalanced Flight Records for Flight Diversion Prediction: A Multi-objective Optimisation Framework

Flight diversions are rare but high-impact events in aviation, making their reliable prediction vital for both safety and operational efficiency. However, their scarcity in historical records impedes the training of machine learning models used to predict them. This study addresses this challenge by proposing a generative augmentation framework for imbalanced aviation tabular records. The principal contribution lies in the design of a composite optimisation objective specifically tailored to flight data, which integrates four complementary quality dimensions into a single score used to guide automated hyperparameter search via the Tree-structured Parzen Estimator (TPE) algorithm: realism, statistical similarity, fidelity, and predictive utility. These dimensions were selected and defined to reflect the operational and statistical requirements specific to aviation records, and were complemented by two descriptive evaluation dimensions, diversity and operational validity, forming a six-stage assessment framework. The composite objective was then used to tune three deep generative models, namely Tabular Variational Autoencoder (TVAE), Conditional Tabular Generative Adversarial Network (CTGAN), and CopulaGAN, with Gaussian Copula (GC) serving as a statistical baseline. Results show that optimised models substantially outperform their default counterparts across all six assessment dimensions, and that augmentation with the resulting synthetic data improves diversion prediction compared to training on real data alone. These findings demonstrate that domain-adapted multi-objective optimisation is an effective strategy for generative augmentation of rare events in aviation, with applicability to other imbalanced tabular prediction tasks.