TailBooster: A Dual-Layer Generative Framework for Extreme Value Augmentation with Operational Validity Enforcement

A recent paper published on arXiv (2608.11951) presents TailBooster, a dual-layer generative framework aimed at enhancing extreme occurrences in air transport data, including significant arrival delays and unusual air times. These infrequent events lead to widespread network disruptions, incurring considerable operational, economic, and safety repercussions. However, their rarity in historical data results in inadequate training signals for machine learning models. Traditional generative models fail to adequately represent distributional tails and do not guarantee operational feasibility, such as pairing a short air time with a long flight distance. TailBooster integrates generative modeling with two anomaly detection layers: a statistical layer identifies extremes using the interquartile range, providing focused training signals to a Tabular Variational Autoencoder, while a deep learning layer ensures operational validity. This framework addresses limitations for mixed-type tabular records, filling a gap not tackled by existing methods. The paper is categorized as a cross-type announcement on arXiv, suggesting submission to multiple categories. Its findings are pertinent to aviation, machine learning, and data augmentation, with implications for enhancing predictive models in air traffic management and safety.