Research map / Graph learning / Knowledge and temporal graphs
Aeolus: A Multi-structural Flight Delay Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Aeolus, a multi-modal flight delay dataset (tabular, flight chains, flight network graphs) from BTS and Meteostat with temporal splits, and benchmarks tabular, sequential, and graph models on it.
Problem
Existing flight delay datasets are small, geographically narrow, flat, leak future information, or are not publicly available, which limits research on delay propagation and generalizable tabular learning.
Approach
Aligns three modalities built from raw flight operations and weather data: a flat tabular table for over 50 million flights, per-aircraft flight chains linked by tail-number over a 24-hour window, and a flight network graph connecting flights via shared airport, aircraft, and crew resources; evaluated with temporal splits.
Key results
- On the tabular benchmark, FTTransformer achieves the best arrival-delay regression MSE (0.914) and TabulaRNN the best classification accuracy (0.772 ARR), with AUCs below 0.63.
- In the flight chain experiments, all sequential models (LSTM, GRU, CNN-LSTM, MogrifierLSTM) reach AUC around 68.5-69.5%, with little variance across architectures.
- Adding VGAE flight-network embeddings to AFM raises AUC from 67.23% to 67.94%, versus 67.65% for DeepWalk.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add VGAE flight-network embeddings to AFM | AFM flight delay classifier · Aeolus ARR delay graph classification | AUC: 0.6723 -> 0.6794 | Helped |
| Add DeepWalk flight-network embeddings to AFM | AFM flight delay classifier · Aeolus ARR delay graph classification | AUC: 0.6723 -> 0.6765 | Helped |
| Replace DeepWalk embeddings with VGAE embeddings | AFM flight delay classifier · Aeolus ARR delay graph classification | AUC: 0.6765 -> 0.6794 | Helped |
| Use random split instead of temporal split (ablation) | Tabular models on Aeolus ARR delay prediction · Aeolus ARR delay prediction | AUC: +0.057 average AUC | Helped |
| Use random split instead of temporal split (ablation) | Tabular models on Aeolus ARR delay prediction · Aeolus ARR delay prediction | ACC: 0.688 -> 0.688; 0.708 -> 0.724; 0.750 -> 0.782; 0.763 -> 0.758; 0.679 -> 0.783; 0.772 -> 0.798; 0.772 -> 0.728 | Mixed |
| Train on moderate disruption instead of severe disruption months (ablation) | COVID-19 ARR delay models · Post-disruption validation | AUC: +0.023 | Helped |
| Train on moderate disruption instead of severe disruption months (ablation) | COVID-19 ARR delay models · Post-disruption validation | ACC: +0.038 | Helped |
| Use FTTransformer rather than the other tested tabular models | Tabular ARR delay regression · Aeolus ARR delay regression | MSE: 0.914 | Helped |
| Use TabulaRNN rather than the other tested tabular models | Tabular delay classification · Aeolus ARR and DEP delay classification | ACC: 0.772/0.761 | Helped |
Surprising findings
- Random splits inflate AUC by an average of 0.057 relative to temporal splits across seven tabular models, indicating temporal leakage.
- Models show high accuracy (0.74-0.77) but low AUC (<0.63) on the tabular benchmark, suggesting class imbalance issues.
Limitations
- Geographic bias: 78.4% of flights from North America and overrepresentation of major hubs.
- Temporal coverage includes COVID-19 disruption (2020-2021), causing non-stationarity.
- Missing operational signals such as ATC decisions, maintenance histories, and passenger flow.
Experimental setup
Datasets: BTS flight records 2016-2024 Meteostat weather data
Benchmarks: Aeolus tabular benchmark Aeolus flight chain benchmark Aeolus flight network benchmark
Metrics: MSE MAE AUC ACC CRPS Recall Precision F1
Baselines: MLP ResNet AutoInt FTTransformer Tangos TabulaRNN SAINT LSTM GRU CNN-LSTM MogrifierLSTM AFM
Builds on: Mambular AFM
Closest papers by meaning
- FlightBERT++: A Non-autoregressive Multi-Horizon Flight Trajectory Prediction Framework (AAAI 2024 · similarity 0.85)
- Effective and Efficient Representation Learning for Flight Trajectories (AAAI 2025 · similarity 0.83)
- NeuralPlane: An Efficiently Parallelizable Platform for Fixed-wing Aircraft Control with Reinforcement Learning (NeurIPS 2024 · similarity 0.81)
- LaKD: Length-agnostic Knowledge Distillation for Trajectory Prediction with Any Length Observations (NeurIPS 2024 · similarity 0.81)
- GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning (ICML 2025 · similarity 0.80)
- AeroGTO: An Efficient Graph-Transformer Operator for Learning Large-Scale Aerodynamics of 3D Vehicle Geometries (AAAI 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.