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

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

TL;DR

This paper introduces a suite of machine learning-based decision support models (DSMs) to predict and explain missed flight connections across different airline operational stages (strategic, pre-tactical, tactical). By leveraging historical flight and passenger data, these models achieve high predictive accuracy (AUC > 0.93) and provide insights into the key factors influencing connection success, offering airlines a data-driven approach to optimize operations and reduce costs.

What Problem Does This Paper Address?

Airlines face a significant challenge in managing connecting flights, where passengers missing their connections can lead to substantial financial losses and diminished customer satisfaction. Traditional approaches often rely on a fixed Minimum Connecting Time (MCT), which may not adequately account for the complex interplay of factors influencing connection success. This paper addresses the need for more sophisticated, data-driven decision support systems that can predict missed connections at various stages of airline operations, from long-term schedule planning to real-time operational adjustments. The authors highlight the lack of research specifically focusing on predicting missed connections using passenger-centric data and ML techniques.

Methods

The core of this work lies in developing four distinct Decision Support Models (DSMs) for different operational horizons: Strategic, Pre-Tactical, Tactical, and Post-Operations. Each DSM employs a machine learning model, primarily XGBoost, to predict the likelihood of a missed connection.

Key methodological contributions include:

  • Probabilistic Encoding and Data Balancing: To handle high-dimensional, heterogeneous, and imbalanced data, the authors employ probabilistic encoding for categorical features and data balancing using Gaussian Mixture Models (GMMs) to oversample the minority class (missed connections).
  • Feature Engineering: New features were engineered, such as "Perceived Connection Time" (scheduled departure minus actual arrival time), which captures the passenger's real-time understanding of their connection window. The "Traffic Network" feature, indicating travel within or outside the Schengen area, was also crucial for accounting for border control procedures.
  • Explainable AI (XAI): SHAP (SHapley Additive exPlanations) values are used to provide post-hoc explanations of the models' predictions, identifying the most influential features and their impact on the outcome. This allows for a deeper understanding of why a connection is predicted to be missed.
  • Multi-stage Modeling: The models are designed to leverage information available at different stages of the flight process, progressively incorporating more data as events unfold.

Datasets

The study utilizes historical data from TAP Air Portugal for the year 2019. This data is comprised of three main sets:

  • Hub Dataset: Contains information about flight arrivals and departures (scheduled times, gates, etc.).
  • Pax Dataset: Includes passenger travel information, such as travel class and connection status.
  • SEF (Border Control) Dataset: Indicates whether a passenger needs to pass through passport control.

This data was integrated and preprocessed, resulting in a dataset of approximately 3.5 million samples for training and evaluation, with about 5.85% representing the minority class (missed flights).

Key Insights and Ideas

  • Scheduled Connection Time is King (Initially): For the strategic planning stage, the scheduled connection time is the most dominant predictor of missed connections, aligning with traditional airline practices.
  • Border Control Matters: The "Traffic Network" feature, which accounts for Schengen area border controls, emerges as a significant predictor, highlighting the impact of these procedures on connection times.
  • Passenger Features Add Nuance, Not Dominance (Yet): While passenger-specific features like age and travel group were introduced in the pre-tactical stage, they did not dramatically improve predictive performance over flight-centric data alone. This suggests that while passenger characteristics might play a role, their influence is less pronounced than timing and logistical factors in this dataset.
  • Real-time Information is Powerful: The Tactical DSM, which incorporates perceived connection time (reflecting actual arrival delays), significantly improves predictive performance, demonstrating the value of real-time operational data.
  • SHAP Reveals Complex Interactions: SHAP analysis reveals that while scheduled connection time is a primary driver, other factors like border control, origin/destination, and even the use of buses for boarding can contribute to missed connections, offering actionable insights for operational adjustments.

Key Results

  • High Predictive Accuracy: All proposed DSMs consistently achieved an Area Under the ROC Curve (AUCROC) higher than 0.93, with the Strategic and Tactical models reaching 0.98 and 0.99 respectively. This represents a significant improvement over baseline models.
  • Cost Reduction Potential: The study indicates that implementing these DSMs for decision-making can lead to substantial cost reductions by enabling preventive measures. The analysis suggests that the cost of preventive actions is often lower than reactive actions when the ratio of reactive to preventive costs exceeds a certain threshold.
  • Improved Feature Importance: SHAP analysis provided clear insights into feature importance across different stages, confirming the dominance of connection time and highlighting the impact of traffic network and, in later stages, perceived connection time.

Limitations and Future Work

  • Seasonality and Evolution: The authors acknowledge that retraining models periodically is necessary due to seasonality and the evolving nature of the aeronautical sector.
  • Out-of-Distribution Behavior: A future line of work suggested is to study the out-of-distribution behavior of the models and adapt them using online learning methodologies.
  • Passenger Data Availability: The paper notes that passenger data is often not publicly available, which can be a limitation for exploring passenger-centric problems.
  • Further Investigation of Passenger Features: While passenger features didn't dramatically boost performance in the pre-tactical stage, the authors suggest further investigation, potentially with more data collection and elicitation of domain knowledge, to better understand their impact.
  • Operational Cost Analysis: The Post-Operations DSM, while valuable for analysis, does not directly support preventive actions. The authors suggest that a more detailed analysis of the effective costs of reactive and preventive actions could further refine the optimal trade-off between precision and recall.

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 AUC of the 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