Extreme events in air transport, such as severe arrival delays and abnormal air times, can trigger cascading network disruptions with heavy operational, economic, and safety costs. However, these events are rare in historical records, leaving machine learning models with insufficient training data. Synthetic data augmentation has been proposed as a solution, but conventional generative models often under-represent distributional tails and may produce operationally infeasible instances, like a short air time paired with a long flight distance.
To address these limitations, researchers at Hugging Face have developed TailBooster, a dual-layer generative framework that combines generative modeling with two anomaly detection layers. The first statistical layer extracts extremes using the interquartile range, providing tail-concentrated training signals to dedicated generative models, specifically a Tabular Variational Autoencoder. The second deep learning layer applies autoencoder-based cleaning to discard synthetic records that violate the operational envelope learned from historical data.
Evaluated on US flight records across five dimensions—diversity, statistical similarity, fidelity, operational validity, and utility—the framework showed marked improvements in operational validity and utility for extreme-event prediction. Across six regression algorithms, training on TailBooster's records reduced Mean Absolute Error by 47-49% for extreme air time and 29-57% for extreme arrival delay prediction compared to conventional synthetic data. Comparable gains were observed when real records were enriched with synthetic extremes.
Being fully data-driven and model-agnostic, TailBooster extends to domains where extreme-event prediction is critical and domain-specific rules are unavailable. The research highlights the potential of combining anomaly detection with generation to synthesize valid extremes, offering a robust solution for rare-event prediction in complex systems.