Abstract
Travel behavior varies across people, shifts with context, and evolves over time. Existing models capture only part of this: discrete choice models analyze one dimension at a time, latent class models fix each traveler to a single class, and hidden Markov models allow switching but ignore episode-level context. This paper proposes a quantum-inspired framework in which a traveler is represented not as a class label but as a dynamic state. In our application, activity type, transport mode, and duration are encoded jointly in a density matrix. Unlike a label, a density matrix holds several behaviors at once, so each profile decomposes by eigenstructure into interpretable sub-patterns. Profiles are learned directly from data, with no outcome variable or class structure imposed. A traveler's state encodes all profiles: context sets the weights before an episode, and the observed behavior updates them after. Habit and adaptation thus emerge from one representation rather than from separate rules. Applied to the Swiss TimeUse+ dataset which consists of 1,310 participants and 146,318 activity episodes, the framework recovers three well-separated behavioral profiles: a Shopper/Leisure profile, a Commuter profile, and a Home Anchor profile. Each profile splits into sub-patterns without supervision, revealing otherwise hidden behaviors, such as car-based versus walk-based shopping trips within a profile. We validate our framework using three established models, Multinomial Logit Model, Joint Latent Class Model, and Hidden Markov Model, along with an ablation study and a simulation, confirming the viability of the framework in counterfactual analysis.This work can be fully reproduced using our open-source codebase on: https://github.com/wissamkontar/A-Quantum-Inspired-Representation-of-Dynamic-Travel-Behavior.