La, Duc Minh (Monash University), Vu, Hai L. (Monash University)

TreeCSP: An Efficient Way to Generate Household Relationships for Population Synthesis

Scheduled for presentation during the Invited Session "S19c-Artificial Transportation Systems and Simulation" (TH-LA-T19), Thursday, November 20, 2025, 16:00−16:20, Surfers Paradise 1

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Population synthesis plays a crucial role in Agent-Based Models (ABMs) and transportation planning by creating realistic representations of individuals and households. However,existing methods often lack flexibility in modeling complex household relationships and integrating domain knowledge without relying on rigid assumptions. Thus, we introduce Tree-based Chained Sampling Pools (TreeCSP), a novel framework that organizes household and person generation through a hierarchical relationship tree and multi-step sampling. TreeCSP enables both fully data-driven learning and the seamless integration of domain knowledge, improving scalability, flexibility, and explainability. Using the Victorian Integrated Survey of Travel and Activity (VISTA) dataset as ground truth, we benchmarked two variants of TreeCSP: CSP-Seed, which uses the VISTA data directly, and CSP-BN, which incorporates Bayesian Networks (BNs), against two widely used methods: Iterative Proportional Updating (IPU) and Conditional Tabular Generative Adversarial Network (CTGAN). Results show that TreeCSP consistently outperforms the baselines, with CSP-BN achieving the lowest Jensen-Shannon Distance (JSD) across both household and relationship attributes. Moreover, TreeCSP effectively captures rare and complex household structures while maintaining model transparency and avoiding infeasible combinations.