TLDR: Aeolus is a large-scale, multi-modal dataset designed to improve flight delay prediction and advance research in tabular data. It uniquely combines over 50 million flight records across three aligned modalities: tabular data with operational and meteorological features, flight chains modeling delay propagation along sequential flight legs, and flight network graphs encoding shared resource connections. This comprehensive dataset supports various machine learning tasks (regression, classification, temporal, graph learning) and addresses limitations of existing datasets by providing robust temporal splits and preventing data leakage, offering a realistic benchmark for both aviation-specific and general structured data challenges.
A new benchmark dataset named Aeolus has been introduced to significantly advance the field of flight delay prediction and support the development of sophisticated machine learning models for tabular data. This dataset addresses critical limitations found in existing resources, which often fail to capture the intricate spatiotemporal dynamics of delay propagation.
Aeolus stands out by offering a multi-structural approach, integrating three distinct yet aligned modalities. Firstly, it provides a comprehensive tabular dataset, featuring rich operational, meteorological, and airport-level details for over 50 million flights. This extensive collection allows for a deep dive into static and contextual factors influencing delays.
Secondly, the dataset includes a flight chain module. This innovative component models how delays propagate along sequential flight legs, effectively capturing upstream and downstream dependencies. Imagine a single aircraft’s journey through multiple flights; a delay in an earlier leg will inevitably affect subsequent ones. The flight chain module helps researchers understand and predict these cascading effects.
Thirdly, Aeolus incorporates a flight network graph. This modality encodes connections based on shared resources like aircraft, crew, and airport infrastructure. By representing these relationships as a network, the dataset enables cross-flight relational reasoning, allowing models to understand how a delay in one flight can impact others through shared resources.
The creators of Aeolus have meticulously constructed the dataset with several key considerations. It features careful temporal splits, comprehensive features, and strict leakage prevention measures to ensure realistic and reproducible machine learning evaluations. This design supports a wide array of tasks, including regression (predicting delay duration), classification (predicting if a delay will occur), temporal structure modeling, and graph learning, making it a unified benchmark across tabular, sequential, and graph data types.
The dataset spans nine years, from 2016 to 2024, covering 320 airports and including over 54 million flight records. It integrates flight data from the U.S. Department of Transportation’s Bureau of Transportation Statistics (BTS) and meteorological data from Meteostat, offering features like airline carrier codes, flight numbers, date components, airport codes, scheduled times, flight durations, and weather measurements (temperature, precipitation, wind speed) at both origin and destination airports.
Aeolus also provides baseline experiments and preprocessing tools to facilitate its adoption by the research community. The dataset is not only valuable for domain-specific flight delay modeling but also for general-purpose structured data research, filling a crucial gap in both areas. For those interested in exploring the dataset and its associated code, more information can be found at the Aeolus research paper.
The research highlights the importance of using temporal splitting strategies over random splits, as random splitting can lead to inflated performance metrics due to temporal leakage. An analysis of the COVID-19 pandemic’s impact on flight delays further demonstrates how exogenous shocks can significantly alter delay patterns, underscoring the need for robust models that can adapt to such shifts.
Also Read:
- Spatiotemporal Error Adjustment Enhances Deep Learning Traffic Models
- Anticipating the Wind: Deep Reinforcement Learning Guides UAVs Through Turbulent City Airflows
In conclusion, Aeolus offers a powerful new tool for researchers and practitioners alike, promising to foster more realistic, multimodal, and generalizable approaches to understanding and predicting flight delays, ultimately contributing to more efficient air traffic management and improved passenger experiences.
