TailBooster: A Dual-Layer Generative Framework for Extreme Value Augmentation with Operational Validity Enforcement
Karim Aly, Alexei Sharpanskykh, Jacco Hoekstra
cs.LG, cs.AI
2026-08-12
TU Delft's TailBooster trains dedicated tabular VAEs on IQR-extracted extremes and autoencoder-filters operationally invalid samples, cutting extreme-event prediction MAE by 29-57% across six regression models on US flight records.
Extreme air transport events, such as severe arrival delays and abnormal air times, make up only a few percent of historical records. Machine learning models get almost no training signal in the tail, so they systematically under-predict the events that matter most. Synthetic data augmentation is the standard fix, but conventional generative models under-represent distributional tails and produce operationally infeasible records, such as a short air time paired with a long flight distance. No prior method addressed both problems for mixed-type tabular data, and existing constraint-based approaches like zGAN require hand-crafted physical rules that must be re-engineered per domain.
TailBooster splits the fix into two independent gates:
The entire validity check is data-driven, with no aviation-specific symbolic rules, so the pipeline transfers to finance or energy domains where governing rules are unavailable. The generative component is model-agnostic.
Evaluated on 60,767 US domestic flights from New York State, January 2023, with all models trained and tested exclusively on extreme subsets. All six regression models improved:
| Training data | Extreme air time MAE | Extreme arrival delay MAE |
| Naive synthetic | 19.40-23.58 min | 34.41-47.58 min |
| Augmented synthetic (TailBooster) | 10.21-12.23 min | 14.89-34.00 min |
| Real data | 6.34-9.88 min | 11.54-32.93 min |
| Real + synthetic extremes | 2.57-8.59 min | 4.49-24.77 min |
That is a 47-49% MAE reduction on extreme air time and 29-57% on extreme arrival delay relative to conventional synthetic data. More practically, enriching real data with synthetic extremes cut delay MAE from 11.54-32.93 to 4.49-24.77 minutes, with Random Forest dropping from 12.09 to 4.49. A classifier's discriminability between real and augmented synthetic records on extreme subsets fell from 0.92/0.88 to 0.54/0.58, indicating the synthetic extremes genuinely resemble the real tail.
For tabular-data practitioners, this is one of the few pipelines that handles tail representation and operational validity together, and both components are lightweight: IQR splitting is a few lines of code, and the cleaning layer is a standard autoencoder. For teams that already hold real data, the message is more direct: mixing in synthetic extremes is a consistent win across all six models, so it is not an artifact of one model family. Code is promised on GitHub upon publication.
The authors list three: evaluation on a single dataset (New York State, January 2023) with unknown cross-season and cross-region generalization; operational validity is assessed visually through correlation plots rather than a quantitative score; and the extreme subsets contain only a few thousand records, limiting the diversity a deep generative model can learn. One caveat from reading the paper: the headline MAE baseline is conventional synthetic data, a baseline already known to under-represent tails, so the 47-49% relative reduction needs that context, and there is no comparison against classical oversampling such as SMOTE.