Aeolus：A Multi-structural Flight Delay Dataset

• This repository provides the full experimental pipeline for building benchmark datasets from the “Aeolus: A Multi‑structural Flight Delay Dataset” (available on Kaggle: mfdd‑multi‑modal‑flight‑delay‑dataset). It demonstrates how to construct the tabular and graph-based datasets from Aeolus, run multimodal baselines, and visualize delay patterns. You can reproduce experiments across modalities and methodologies, and explore visualizations that help analyze and understand flight delay dynamics.

---

📂 Repository Structure

• Datasets/: Data extraction, tabular and graph constructors.
• exp/: Experimental scripts for chain, network, and tabular baselines.
• util/: Utility modules for validation and figure generation.

---

📊 Dataset Schema & Features

Feature Name	Description	Example
FL_DATE	Flight date/time	2024/01/01 00:00
OP_CARRIER	Operating airline code	9E
OP_CARRIER_FL_NUM	Flight number	4814
ORIGIN / DEST	Origin and destination airport IATA code	JFK / DTW
CRS_DEP_TIME	Scheduled departure time	12:52
DEP_TIME, DEP_DELAY	Actual departure time and delay in minutes	12:47, -5
ARR_TIME, ARR_DELAY	Actual arrival time and delay	14:49, -19
TAXI_OUT, TAXI_IN	Taxi-out and taxi-in durations	31, 7
CRS_ELAPSED_TIME, ACTUAL_ELAPSED_TIME, AIR_TIME	Flight durations	136, 122, 84
MONTH, DAY_OF_MONTH, DAY_OF_WEEK	Calendar features	1, 1, 1
ORIGIN_INDEX, DEST_INDEX	Airport index for embedding lookup	166, 93
O_TEMP, O_PRCP, O_WSPD	Origin weather: temperature, precipitation, wind speed	4.4°C, 0 in, 11.2 mph
D_TEMP, D_PRCP, D_WSPD	Destination weather	0°C, 0 in, 13 mph
O_LATITUDE, O_LONGITUDE, D_LATITUDE, D_LONGITUDE	Geographic coordinates	40.63975, -73.77893

---

🎯 Key Capabilities

• Tabular data construction for ARR_DELAY and DEP_DELAY tasks
• Flight-chain modeling (aircraft & crew continuity)
• Flight-graph construction (airport/airspace resource interactions)
• Support baseline models: MLP, AutoInt, TabulaRNN, ResNet, VGAE+AFM， etc.

---

🔬 Experiments & Visualization

• Tabular baseline experiments under exp/Tab_exp
• Time-based modeling pipelines under exp/Chain_exp
• Graph-based modeling pipelines under exp/Network_exp
• Some visualizations are shown below. For more visualizations, see util/figures

SHAP Summary Plot

• The SHAP summary beeswarm plot illustrates how the top‑15 features influence model predictions—each point shows an instance’s feature value and its positive or negative contribution (SHAP value) to delay prediction, with color encoding feature magnitude and horizontal position indicating impact on output

Delay propagation visualization in Flight Chains

• The delay propagation visualization plots a flight chain across hub airports—with color‑coded lines showing average arrival delays—highlighting how an upstream delay travels through connections to downstream locations.

Delay propagation visualization in Flight Networks

• The delay propagation visualization highlights how average arrival delays flow through connected airports in the flight network—edges colored by delay indicate how upstream disruptions spread across routes in the airport graph.

Monthly Average Arrival Delay Trends (2019-2022) with COVID-19 Impact

• This graph illustrates the monthly average arrival delay for the years 2019 to 2022, highlighting the atypical delay patterns during the COVID-19 pandemic, particularly between March and June 2020, when flight disruptions caused an abnormal trend in the data.

---

🚧 Limitations & Future Work

• Lacking fine-grained operational signals such as real‑time ATC interventions, crew rotation, or passenger itineraries
• Strong geographic bias: ~78% of flights originate in North America; limited coverage for regional hubs in Africa or South Asia
• Large scale and multimodal nature may impose high computational requirements; reproducibility supported via full configuration scripts, with lightweight subsets planned