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

The Problem: Rare Extremes, Unrealistic Synthetics

Machine learning models trained on historical data struggle with rare but critical events. In aviation, severe delays and unusual flight times are infrequent—yet their cascading effects across networks are costly and dangerous. The challenge is twofold: standard datasets don't contain enough extreme examples to train predictive models, and when generative models synthesize new data to fill this gap, they either ignore the tails of distributions or produce operationally impossible records (like a 15-hour flight time for a 500-mile route). No existing method simultaneously solves both problems for mixed-type data (numbers, categories, etc.).

How TailBooster Works

The framework operates in two stages, each adding a layer of validity.

First stage: statistical extraction and targeted generation. TailBooster identifies extreme values using the interquartile range—a simple, robust statistical measure. Records falling in these extremes become the training signal for a Tabular Variational Autoencoder (a neural network that learns compressed representations of data and can generate new samples). By feeding this model only tail-concentrated data, it learns to generate synthetic extremes rather than averaging them away, which is what conventional generative models do.

Second stage: operational cleaning. The synthetic records are then passed through a deep learning filter—an autoencoder trained on historical data. This filter learns the "operational envelope," the set of physically and operationally valid combinations of variables. Any synthetic record that violates this learned envelope (e.g., impossible speed-distance pairs) is discarded. This cleanup is fully data-driven; no hand-coded rules are needed.

The authors evaluated their framework on US flight records across six dimensions: diversity (do generated records vary enough?), statistical similarity (do they match real distributions?), fidelity (are they faithful to the training data?), operational validity (are they physically plausible?), and utility (do they actually improve downstream prediction?).

Why It Matters

Results show substantial gains. When predicting extreme air times, models trained on TailBooster's synthetic data achieved 47–49% lower error compared to conventional synthetic data; for extreme arrival delays, 29–57% improvement. The operational cleaning layer was critical—it sharply reduced invalid records while maintaining statistical properties.

Crucially, the framework is model-agnostic and requires no domain expertise. Unlike aviation-specific rule sets, TailBooster learns validity constraints directly from data, making it portable to other critical domains (power grids, infrastructure, logistics) where extreme-event prediction matters and historical extremes are scarce.

The authors also showed that enriching real records with synthetic extremes yields comparable gains, suggesting a practical hybrid approach when data is partially available.