Decision Support Models for Predicting and Explaining Airport Passenger Connectivity From Data

Marta Guimarães, Cláudia Soares, Rodrigo M. M. Ventura

Introduction

Missed flight connections represent a significant logistical and financial burden for the aviation industry. Beyond the direct "reactive costs"—such as rebooking fees, hotel vouchers, and meal expenses—airlines suffer a long-term loss of customer goodwill. Traditionally, airlines have managed these risks using a static Minimum Connection Time (MCT). This "one-size-fits-all" rule is often too blunt, failing to account for the dynamic nature of airport operations, including arrival delays, passenger demographics, and specific terminal bottlenecks like border control.

This paper introduces a suite of machine learning-based Decision Support Models (DSMs) designed to replace static MCT rules with dynamic, data-driven predictions. By utilizing XGBoost and SHAP (Shapley Additive Explanations), the researchers provide a framework that not only predicts missed connections with high accuracy but also explains the underlying reasons for each prediction, enabling proactive operational management.

Methodology: A Multi-Stage Approach

The researchers developed four distinct models tailored to different operational horizons, allowing for decision support ranging from long-term scheduling to real-time gate management.

The four operational stages of the proposed Decision Support Models: Strategic, Pre-Tactical, Tactical, and Post-Operations.
  1. Strategic: Long-term planning based on flight designators and schedules.
  2. Pre-Tactical: Short-term planning once passenger data (e.g., age, travel class) becomes available after ticket purchase.
  3. Tactical: Real-time support incorporating actual arrival times and "perceived" connection windows.
  4. Post-Operations: Retrospective analysis to evaluate past performance and refine future models.

The core of these models is the XGBoost algorithm, an ensemble of gradient-boosted decision trees known for high performance on tabular data. To address the class imbalance inherent in missed connection data (where missed connections are the minority class), the team employed Gaussian Mixture Models (GMM) for oversampling. The study was validated using a massive real-world dataset of ~3.5 million passenger samples from TAP Air Portugal’s Lisbon (LIS) hub.

Explainability and Key Insights

A critical component of this research is the move away from "black box" AI. By integrating SHAP, the models provide human-interpretable explanations for their outputs, allowing airline staff to understand why a specific passenger is flagged as "at risk."

SHAP summary plot illustrating the impact of various features on the model's output. Scheduled Connection Time and Traffic Network are among the most influential factors.

Analysis of the model's features revealed several key operational insights:

  • The "Bus" Penalty: Flights serviced by buses rather than jet bridges carry a higher risk of missed connections, as boarding must conclude earlier to allow for transit time to the aircraft.
  • Traffic Network Bottlenecks: Connections requiring movement between Schengen and non-Schengen zones (necessitating passport control) were identified as high-risk factors.
  • Demographic Proxies: While age was a significant predictor—likely reflecting mobility differences through the terminal—gender had almost no impact on predictive power.
  • Perceived Connection Time: A novel feature defined as the scheduled departure time minus the actual arrival time of the incoming flight proved highly effective in the Tactical phase.

Performance and Economic Impact

The proposed models significantly outperformed the industry-standard MCT baseline across all metrics:

  • Accuracy: The Strategic and Pre-Tactical models achieved an AUC-ROC of 0.98, while the Tactical model reached 0.99.
  • Precision-Recall: In the Strategic phase, the model achieved an AUC-PR of 0.90, compared to just 0.26 for the baseline MCT-based approach.

From an economic perspective, the researchers demonstrated that these high-precision models enable "preventive" actions (e.g., short flight delays or passenger assistance) that are far more cost-effective than "reactive" fixes. The models remain economically viable as long as reactive costs are at least 1.16 to 1.37 times higher than preventive costs—a threshold easily met in standard airline operations.

Limitations and Future Directions

While highly effective, the models require periodic retraining to account for seasonal shifts in travel patterns (e.g., summer holiday rushes vs. winter weather). Future work will explore "online learning" systems to update models incrementally as new data flows in. Additionally, the current framework does not explicitly account for "black swan" events, such as global pandemics or major strikes, which fall outside historical data distributions.

Summary updated

Abstract

Predicting if passengers in a connecting flight will lose their connection is paramount for airline profitability. We present novel machine learning-based decision support models for the different stages of connection flight management, namely for strategic, pre-tactical, tactical and post-operations. We predict missed flight connections in an airline’s hub airport using historical data on flights and passengers, and analyse the factors that contribute additively to the predicted outcome for each decision horizon. Our data is high-dimensional, heterogeneous, imbalanced and noisy, and does not inform about passenger arrival/departure transit time. We employ probabilistic encoding of categorical classes, data balancing with Gaussian Mixture Models, and boosting. For all planning horizons, our models attain an area under the curve (AUC) of the receiver operating characteristic (ROC) higher than 0.93. SHAP value explanations of our models indicate that scheduled/perceived connection times contribute the most to the prediction, followed by passenger age and whether border controls are required.
BibTeX
@article{guimaraes2022decision,
  author = {Marta Guimarães and Cláudia Soares and Rodrigo M. M. Ventura},
  title = {{Decision Support Models for Predicting and Explaining Airport Passenger Connectivity From Data}},
  journal = {IEEE Trans. Intell. Transp. Syst.},
  year = {2022},
  doi = {10.1109/tits.2022.3147155},
}
Figure