Machine learning models for reliable airline ancillary pricing

Machine learning is becoming increasingly prevalent for decision-making across key application areas such as healthcare, finance, law systems, and pricing. However, evaluating the predictive power of models on historical data is not enough. When deploying ML models in the real-world, system designers may wish that models exhibit certain properties which encourage trust, interpretability, and robustness. In this study, we discuss two such properties and demonstrate their relevance through experiments on an airline ancillary use-case. From a qualitative lens, designers may want the trained model to exhibit particular shape behavior that conforms to prior domain knowledge. Trends such as monotonicity, convexity, diminishing or accelerating returns are some of the desired shapes. Conformance to these shapes makes the model more interpretable for system designers, and adequately fair for customers. We notice that many such common shapes are related to derivatives, and propose a new approach, PenDer (Penalizing Derivatives), which incorporates these shape constraints by penalizing the derivatives. We further present an Augmented Lagrangian Method (ALM) to solve this constrained optimization problem. Experiments on three real-world datasets illustrate that even though both PenDer and state-of-the-art Lattice models achieve similar conformance to shape, PenDer captures better sensitivity of prediction with respect to intended features. We also demonstrate that PenDer achieves better test performance than Lattice while enforcing more desirable shape behavior. For the airline, PenDer appropriately balances the trade-off between constraint satisfaction and predictive performance -- leading to potential increased revenues compared to other models. Another important expectation of ML models is generalization to unseen data, that is, performance on new data. Textbook approaches assume that the data distribution at test-time is similar to the training distribution. In practice, however, this could be violated because of ever-changing customer behavior or a dynamic environment. Motivated by airlines' concerns of model performance drop during the COVID-19 pandemic, we study covariate shift to examine changes in the underlying data. We focus on shift detection before and during COVID-19 through a blend of (i) discriminative model training to distinguish train from test, and (ii) statistical testing of estimated feature densities. After experiments, we note that customer behavior has changed with notable changes being an increase in advance bookings, one-way trips, and individual travel compared to group travel. Presence of such drift indicates that the airline must perform correction before deployment for optimal performance.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Industrial Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2021

Author dc:creator

• Gupta, Akhil

Contributors dc:contributor

• Marla, Lavanya

• Machine Learning
• Pricing
• Airline Ancillary
• Neural Networks

Statement dc:rights

• Copyright 2021 Akhil Gupta

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/110643

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/110643