Student(s): Kai Puma-Stehlik, Sondo Yoon
Project: Multi-Task Learning for Predicting Customer Behavior in the Airline Industry
Advisor(s): Meng Zhao
Abstract
Accurately predicting passenger intent, booking timing, and upsell likelihood directly impacts airline revenue and personalization, yet most models ignore sequential patterns underlying traveler behavior. This work addresses that gap by reframing passenger prediction as a unified, time-dependent problem rather than a set of isolated tasks.
We present a Temporal Tri-Task Multi-Task Learning (MTL) framework that jointly predicts three targets from a single shared representation: trip purpose classification, advance purchase regression (days between booking and departure), and upsell/ancillary propensity. While prior work addresses these targets in isolation, our architecture captures their interdependencies end-to-end.
Two novel contributions distinguish this work from existing airline MTL literature. First, a GRU-based temporal encoder operates over each passenger’s booking history, constructing a context vector that summarizes prior travel patterns before fusing it with target-booking features in the shared trunk. This treats booking sequences as temporally ordered data rather than independent observations, allowing the model to exploit behavioral consistency signals that cross-sectional models discard. Second, we extend the standard two-task MTL formulation by adding an upsell prediction head, directly linking intent modeling to revenue conversion. GradNorm dynamically reweights task-specific gradients each training step, preventing any single objective from dominating the shared representation. Temporal integrity is enforced via Purged GroupKFold cross-validation with an embargo gap to eliminate leakage at fold boundaries.
Applied to 200,000+ bookings, our baseline bi-task MTL achieves F1=0.60 for trip intent and R²=0.90 for lead time with MAE under five days. The tri-task temporal model expects the largest gains on upsell propensity, where prior ancillary behavior provides the strongest predictive signal.
About Kai Puma-Stehlik
Major: Industrial and Systems Engineering
Coming Soon
About Sondo Yoon
Major: Industrial and Systems Engineering
Coming Soon