Predicting Spatial Choice Without Historical Data: A Theory-Guided LLM Simulation Framework

Lee, S., Y. Yu, & D. Shin

In preparation for Journal of Management Information Systems

Working Paper

Overview

Predicting how people will use a space that does not yet exist is a cold-start problem: there is no usage history to learn from, and stated preferences collected through surveys often diverge from what people actually do. This paper asks whether a language model can stand in for the data a new market does not have, provided its reasoning is constrained by behavioral theory rather than by history.

We develop the Person-Environment-Situation (P.E.S.) framework, which structures LLM reasoning around Lewin's field theory, and validate it against real facility usage records with the later period held out. The result is not that a language model imitates people, but that theory can substitute for the observational data a novel setting lacks.

Motivation

The cold-start problem in facility design

Design decisions for new buildings carry long-term economic and social consequences, yet the behavior they aim to accommodate cannot be observed in advance. Existing approaches fall short in predictable ways:

Surveys capture stated preferences that diverge from actual usage
Observational studies require an existing facility to observe
Space Syntax and layout planning models capture spatial logic but not behavioral complexity
Supervised prediction models require the very usage history that does not yet exist

Why theory rather than history

If a model cannot learn from past behavior in this setting, its predictions must be constrained by something else. We use behavioral theory as that constraint, which also makes the reasoning behind each prediction inspectable rather than opaque.

The P.E.S. Framework

Kurt Lewin's field theory holds that behavior is a function of the person and the environment, B = f(P, E). The framework operationalizes this for language model reasoning by separating three inputs that are specified independently for each simulated decision.

Person

Demographic attributes and household composition
Preferences, goals, and current intentions
Constraints that shape what options are available to this person

Environment

Physical layout and accessibility of the facility
Available amenities and what each affords
Ambient conditions and the presence of others

Situation

Time of day, day of week, and season
Purpose of the visit and competing obligations
Social context in which the choice is made

Validation

Data

Real facility usage records from residential complexes, spanning 18 months
The later period is held out entirely, so the model never sees the behavior it predicts
Predictions are compared against observed usage distributions rather than point outcomes

Comparison

Supervised learning baselines trained on historical usage
Existing LLM-based simulation tools
Distributional agreement measured by Jensen-Shannon Divergence

Related Output

PlaceSim (CIKM 2025)

An earlier version of this framework was released as PlaceSim, a web-based platform that lets architects, planners, and facility managers run these simulations without writing code. It was published in the Proceedings of the ACM International Conference on Information and Knowledge Management. Read more · doi:10.1145/3746252.3761461

Keywords

LLM-based Behavioral Simulation Cold-Start Prediction Discrete Choice Environmental Psychology Facility Design Spatial Analytics