PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Karim Aly, Alexei Sharpanskykh, Jacco Hoekstra

arXiv:2608.11951v1cs.LGcs.AI

TL;DR

Extreme air-transport events are rare, while conventional generators under-represent distributional tails and may produce operationally infeasible records. TailBooster combines tail-focused augmentation with data-driven operational cleaning, reducing extreme-event prediction error across six regression models by 47–49% for air time and 29–57% for arrival delay versus conventional synthetic data.

  • Problem

    Conventional tabular generators under-represent distributional tails and cannot ensure operationally feasible multivariate records, limiting extreme-event augmentation.

  • Method

    TailBooster combines statistical tail extraction, a TVAE generative stage, and deep-learning anomaly detection to generate and operationally filter mixed-type flight records.

  • Results

    47–49% and 29–57% MAE reductions were achieved for extreme air time and arrival delay, respectively, versus conventional synthetic data across six regression models.

  • Takeaways & Limitations

    Targeted augmentation and data-driven cleaning improved the predictability of extreme events while reducing operationally implausible synthetic records.

  • Takeaways & Limitations

    Domain-constrained generative approaches generally require known, formalisable governing laws, limiting transferability across operational settings.

Abstract

from arXiv · show

Extreme events in air transport, such as severe arrival delays and abnormal air times, cause cascading network disruptions with substantial operational, economic, and safety costs. Such events are rare in historical records, leaving insufficient training signal for machine learning models. Synthetic data augmentation offers a principled solution, but conventional generative models under-represent distributional tails and give no guarantee against operationally infeasible instances, such as a short air time paired with a long flight distance. No existing approach addresses both limitations for mixed-type tabular records. We propose TailBooster, a dual-layer generative framework combining generative modelling with two anomaly detection layers. A statistical layer extracts extremes via the interquartile range, supplying tail-concentrated training signal to dedicated generative models, here a Tabular Variational Autoencoder. A deep learning layer then applies autoencoder-based cleaning, discarding synthetic records that violate the operational envelope learned from historical data. The framework was evaluated on US flight records across five dimensions: diversity, statistical similarity, fidelity, operational validity, and utility, the latter two being the primary improvement targets. Data-driven cleaning markedly improved operational validity, while targeted augmentation enhanced utility for extreme-event prediction. Across six regression algorithms, training on the framework's records reduced Mean Absolute Error by 47-49% on extreme air time and 29-57% on extreme arrival delay prediction relative to conventional synthetic data, with comparable gains 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.

1. Introduction

TailBooster addresses sparse extreme-event training signal in flight records by combining generative modelling with anomaly detection to improve tail representativeness and operational validity. The study evaluates whether targeted synthetic augmentation improves regression accuracy for extreme air times and arrival delays.

  • Motivation: Extreme arrival delays and abnormal air times can trigger cascading disruptions across air-transport networks, imposing substantial operational, economic, and safety consequences.These consequences affect airlines, airports, passengers, and air traffic management.
  • Problem: Rare, sparsely populated distributional tails provide insufficient training signal, causing regression models to perform poorly on extreme operational outcomes.Air time and arrival delay are continuous metrics, so their prediction is formulated as regression.
  • Motivation: Synthetic data can enrich tail regions for practitioners with real records and support model development when historical flight data are inaccessible.The introduction identifies constrained data sharing and limited public accessibility as additional obstacles.
  • Method: TailBooster combines a TVAE generative stage with statistical pre-generation and deep-learning post-generation anomaly detection layers.The framework targets extreme representativeness and operational validity in mixed-type tabular flight records.
  • Contributions: The study reports improved operational validity from data-driven cleaning and improved extreme-event predictability from targeted augmentation across six regression models.It evaluates gains for both synthetic records and augmentation of real historical records.

2. Related work

Related work spans tabular generative architectures, EVT-based tail characterisation, EVT-augmented generators, and operational-validity approaches. It identifies persistent gaps in tail representation, mixed-type tabular synthesis, regression utility for extremes, and domain-independent validity enforcement.

  • Tabular generative models: GANs, VAEs including TVAE, and diffusion models dominate tabular data synthesis, but their objectives prioritise high-density regions and under-represent distributional tails.TVAE is described as less susceptible to GAN mode collapse and computationally lighter than diffusion models.
  • Extreme Value Theory: EVT characterises distributional tails, but it is a single-feature tail estimator rather than a generator preserving multivariate dependencies, heterogeneous data types, and operational constraints.The reviewed foundation includes Block Maxima and Peaks-Over-Threshold methods, with POT generally preferred for sporadic continuous extremes.
  • EVT-augmented generation: EVT-augmented models modify training objectives, sampling or conditioning, or latent representations, yet primarily target continuous homogeneous data and therefore do not address mixed-type tabular records.Examples include Pareto GAN, EV-GAN, ExGAN, evtGAN, HTGAN, and ExtVAE.
  • Existing extreme-tabular frameworks: Prior outlier-focused frameworks emphasise binary-classification utility measured by Area Under the Curve rather than regression accuracy on extreme targets, and lack explicit operational-validity verification.The cited limitation also notes dependence on GAN architecture and covariance geometry for outlier magnitude.
  • Operational validity: Domain-informed generators can produce operationally credible extremes, but their physical laws and constraints are domain-specific, limiting transferability when governing rules are unavailable.Examples include mass conservation for flood records and grid-stability conditions for power-grid scenarios.

3. Methodology

TailBooster’s methodology uses U.S. flight records, established preprocessing and feature engineering, and a sequential framework combining generative modelling with statistical and learned anomaly detection. Ten features drive generation, while relational features are derived afterward to preserve dependencies and support realistic synthetic records.

  • Data sources: Flight records came from the publicly accessible TranStats Database for Airline On-Time Performance maintained by the Bureau of Transportation Statistics.The database contains U.S. domestic flight information on operational delay types, cancellations, and diversions.
  • Preprocessing: The study applied the established preprocessing and feature engineering procedure to all January 2023 arrivals and departures in New York State, excluding cancelled and diverted flights.The target metrics include the continuous variables “Air Time (min)” and “Arrival ∆T (min).”
  • Feature engineering: Ten features were selected as direct generative-model inputs, comprising six temporal and four categorical variables.Remaining attributes were designated relational features and derived after generation to preserve inter-feature dependencies and enhance synthetic-record realism.
  • Framework architecture: The TailBooster pipeline sequentially combines two anomaly detection layers, a generative stage, and a relational validity filter.The anomaly detection layers comprise one statistical layer and one learned from historical data.

A. Input

TailBooster uses the full historical flight dataset together with user-defined target features and operationally correlated features. In the study, the targets were Air Time and Arrival ΔT.

  • A. Input: The framework takes the full historical dataset, denoted D = {x(i)}N, as its primary input.Each x(i) denotes a single flight record.
  • A. Input: Users specify target features T containing continuous operational metrics whose distributional tails are of interest.The target list determines which operational metrics receive tail-focused treatment.
  • A. Input: Users also provide operationally correlated feature names Xc to train the autoencoder-based anomaly-cleaning layer.These features are extracted from historical data and support operational validity enforcement.
  • A. Input: In the study, T = {“Air Time (min)”, “Arrival ΔT (min)”}, giving Ntf = 2.The historical dataset consisted of U.S. domestic flight records.

B. IQR-based extreme subset extraction

TailBooster’s first anomaly-detection layer extracts target-feature extremes using Tukey fences based on the interquartile range. Applied independently to each target, this produces one extreme subset per feature alongside the full historical dataset for generative-model training.

  • Extreme subset extraction: The first anomaly-detection layer partitions records according to the distributional tails of each target feature.The procedure operates independently for every target feature.
  • Extreme subset extraction: The extraction yields Ntf + 1 datasets: the full historical record plus one extreme subset for each target feature.In this study, the collection comprises three datasets in total.
  • Extreme subset extraction: Air Time (min) extremes lie outside [−70.00, 282.00] minutes, while Arrival ∆T (min) extremes lie outside [−65.50, 58.50] minutes.These study-specific intervals define the extreme records for the two target features.
  • Extreme subset extraction: The IQR criterion isolates distributional tails without parametric assumptions and is suited to mixed-type, heavy-tailed operational flight records.The resulting datasets are subsequently used to train the generative models.

C. Generative model training

TailBooster trains separate generative models on the full historical record and each target-feature extreme subset, using TVAE for mixed-type tabular data. Independently sampled outputs are generated at a 1.2 ratio to offset records expected to be removed during post-generation filtering.

  • Model training: One model is trained on the full historical record, while separate models train on each extreme subset to amplify target-feature tail-region signal.The full-record model preserves nominal patterns; extreme-subset models focus training on corresponding tails.
  • Model training: TVAE models mixed-type tabular records with a probabilistic encoder and decoder that reconstructs records from sampled continuous latent vectors.Its objective combines reconstruction with KL-divergence regularisation toward a prior, supporting sampling from a structured latent space.
  • Model training: TVAE was selected for training stability, reliable convergence, reduced mode collapse, and lower computational cost than diffusion-based models.The framework remains generative-model-agnostic, allowing substitution by a conditional GAN or another tabular generative model.
  • Synthetic sampling: 1.2 was the sampling ratio, compensating for records expected to be removed during subsequent post-generation filtering stages.Records are sampled independently from all trained generative models to form the synthetic dataset.

D. Relational validity filter

The relational validity filter rejects synthetic records with origin–destination airport pairs absent from historical data. The remaining validity-filtered datasets are passed to the operational cleaning layer, with the uncleaned full-record output designated as the Naïve Synthetic baseline.

  • Relational validity filter: Synthetic records are rejected when their origin–destination airport pair is absent from the historical records.The historical route set is defined as PD = {(o(i), d(i)) : x(i) ∈ D}.
  • Relational validity filter: A synthetic record is accepted if and only if its origin–destination pair belongs to PD, ensuring routes observed in historical data remain.Records satisfying this condition are operationally plausible routes observed historically.
  • Relational validity filter: Applying the filter independently to each ˜S_k produces validity-filtered datasets that proceed to the operational cleaning layer.The resulting datasets are denoted ˜Svalid_k for k = 0, . . . , Ntf.
  • Relational validity filter: ˜Svalid_0 is designated Snaïve, the Naïve Synthetic dataset and baseline output of conventional generation without operational cleaning or targeted extreme value augmentation.It is trained on the full historical record.

E. Autoencoder-based operational cleaning

TailBooster’s second anomaly-detection layer uses autoencoders trained on historical records to learn operational feature correlations and remove synthetic records that violate those learned norms. A conservative 99th-percentile reconstruction-error threshold filters implausible records while retaining most operationally plausible samples.

  • E. Autoencoder-based operational cleaning: Autoencoders are trained once per dataset on historical records, then applied to remove synthetic records violating empirically learned operational norms.The process separates model training from operational cleaning.
  • E. Autoencoder-based operational cleaning: Joint reconstruction learning internalises inter-feature constraints, such as proportionality between route distance and air time, without symbolic domain rules.These learned correlations characterize valid flight operations within each training distribution.
  • E. Autoencoder-based operational cleaning: 99th-percentile reconstruction-error thresholds provide conservative anomaly cutoffs designed to minimise discarding operationally valid synthetic records.For each dataset, the threshold is computed from reconstruction errors on historical operational features.
  • E. Autoencoder-based operational cleaning: Synthetic records are discarded when reconstruction error exceeds the dataset-specific threshold, producing cleaned datasets for subsequent output construction.The threshold retains most operationally plausible records while removing physically inconsistent flights.

F. Output

TailBooster produces three evaluation datasets alongside the original Real dataset reference: Naïve Synthetic, Augmented Synthetic, and Augmented Real. These datasets support comparisons of extreme augmentation, operational cleaning, and downstream predictive performance, with generated records also checked for structural integrity.

  • Dataset outputs: TailBooster produces three datasets—Snaïve, Saug, and Daug—for evaluating the framework’s usefulness.Snaïve is the validity-filtered output of a model trained on the full historical record, while Saug and Daug incorporate cleaned synthetic extremes.
  • Dataset outputs: Snaïve is the conventional synthetic-data baseline against which TailBooster improvements are assessed.It excludes operational cleaning and targeted extreme-value augmentation.
  • Dataset outputs: Saug merges synthetic records that survive both validity filtering and the second anomaly-detection layer, making it TailBooster’s primary output.Its comparison with Snaïve isolates the improvement achieved over conventional synthetic generation.
  • Dataset outputs: Daug combines cleaned synthetic extreme records with real historical data to measure the utility gain from enriching the record with operationally valid extremes.The synthetic extremes are generated from models trained on target-specific extreme subsets and must survive both filtering stages.
  • Validation checks: Generated datasets undergo structural integrity checks for feature-count preservation, observed continuous-variable ranges, and valid discrete categories.These checks establish credibility before downstream analysis.

4. Results

TailBooster preserved or improved synthetic-data quality while substantially improving coverage and validity for extreme air times and arrival delays. These gains translated into markedly better extreme-event prediction than conventional generation, without meaningful evidence of memorisation.

  • Data quality: TailBooster preserved diversity and improved overall statistical similarity from 79.98% for Naïve Synthetic to 86.26% for Augmented Synthetic.The PCA projections showed overlapping real-data clusters, with modest diversity gains for Augmented Synthetic.
  • Data quality: 0.78 overall discriminability for Real vs. Saug was lower than 0.88 for Real vs. Snaïve, indicating greater similarity to real data.The reduction reflected removal of 3,091 anomalous records and the addition of higher-fidelity, operationally valid extreme records.
  • Data quality: Extreme-subset discriminability fell from 0.92 to 0.54 for Air Time and from 0.88 to 0.58 for Arrival ∆T when using Saug.The results indicate that targeted extreme-record generation made synthetic extremes more similar to real extremes.
  • Data quality: Only one of 14 Naïve Synthetic records flagged as potential copies was an exact duplicate, corresponding to 0.002% of the dataset.Manual inspection classified the other 13 flagged records as generalisation rather than memorisation.
  • Operational validity: TailBooster covered extreme long-haul air times and removed implausible air time–distance pairs through augmentation and autoencoder-based operational cleaning.Conventional generation failed to cover observed extreme air times and arrival delays while producing operationally implausible records.
  • Prediction utility: 47–49% lower MAE was achieved for extreme Air Time, while extreme Arrival ∆T MAE decreased by 29–57% relative to Naïve Synthetic across six models.Augmented Real also outperformed Real alone for extreme Arrival ∆T prediction.

5. Discussion

The discussion favors retaining separate airport identifiers during generation and enforcing route validity afterward, because fixed route encoding can collapse operational coverage. It also emphasizes TailBooster’s practical transferability, model-agnostic design, computational trade-offs, and study limitations.

  • Route-validity design: FixedCombinations can prevent unseen origin–destination pairs but collapses two airport features into one route token, limiting airport-specific operational and temporal associations.The constraint reduces broken relational structure by encoding each observed airport pair as a single categorical identifier.
  • Route-validity design: Post-generation route filtering preserves separate airport inputs and provides more uniform operational-space coverage than FixedCombinations without distributional collapse.Figure 6 contrasts collapse under FixedCombinations with uniform coverage from the post-generation validity filter.
  • Practical implications: TailBooster enriches historical datasets for airlines, airports, and air navigation providers by augmenting under-represented operational tails to improve extreme-event predictability.The discussion identifies severe arrival delays and abnormal air times as examples of operationally critical extremes.
  • Transferability and adoption: The framework transfers beyond aviation because it learns empirical inter-feature correlations from historical records rather than relying on hand-crafted domain rules.Its model-agnostic pipeline also permits substituting the TVAE without changing other components.
  • Computational considerations: 60,767 flight records and 100 hyperparameter-optimisation trials per generative model dominate computation, while tail-subset models add negligible overhead.The extreme subsets contain 3,726 records for E(1) and 5,470 for E(2).
  • Limitations: The study is limited to U.S. domestic New York State flights from January 2023, and operational validity is assessed visually through pairwise correlation plots.Generalisation across seasons, regions, and airport network structures remains to be established.

6. Conclusions & Future Work

TailBooster addresses tail under-representation and operational invalidity in mixed-type aviation data while preserving or improving conventional generation qualities. Its data-driven, model-agnostic design supports extreme-value augmentation across settings, with future work focused on broader evaluation and sensitivity to fixed design choices.

  • Conclusions: TailBooster combines IQR-based extreme-subset extraction, dedicated TVAE models, and autoencoder-based anomaly detection to target distributional-tail under-representation and operationally invalid synthetic records.The framework is designed for mixed-type tabular aviation records and addresses two complementary failure modes of conventional deep generative models.
  • Conclusions: Diversity was maintained across real clusters, while statistical similarity and fidelity improved relative to the Naïve Synthetic baseline.These preservation checks indicated that augmentation and cleaning did not degrade qualities achieved by conventional generation.
  • Conclusions: Operational cleaning markedly reduced operationally implausible synthetic records, meeting one of TailBooster’s two primary improvement targets.The passage identifies operational validity and utility as the primary improvement targets, although it does not provide their numerical results here.
  • Conclusions: TailBooster benefits users with historical flight records by augmenting real data with operationally valid synthetic extremes and users without such records through improved synthetic alternatives.The framework is fully data-driven and generative-model-agnostic, supporting extension to other domains.
  • Future Work: Future work should evaluate TailBooster across multiple months, more airports, and more airline operators, and test sensitivity to the fixed IQR multiplier for continuous targets.Embedding domain-specific equations could increase precision but would require re-engineering for each operational setting and reduce transferability.
Loading 2608.11951v1…