WGAN Boosts Synthetic Population Diversity

• Architecture: Three-component WGAN-GP model consisting of generator and two critics designed to handle different parts of generated data • Optimization: Optimizes Wasserstein distance between real and generated data distributions with gradient penalty to enforce 1-Lipschitz constraint for stable training • Regularization: Inverse Gradient Penalty (IGP) term added to generator loss function to address mode collapse and improve sample diversity • Data Integration: Fuses complementary datasets—census data (comprehensive socio-demographic attributes) with travel survey data (rich mobility information but limited coverage and sample bias) • Problem Formulation: Population synthesis operates on individual-level survey data in tabular form where each row represents a population agent with multiple attributes • Evaluation Framework: Unified metric for similarity assessment with special emphasis on recall, precision, and F1 score for diversity and feasibility measurement[1][3]

This advancement addresses a critical bottleneck in agent-based modeling for transportation and urban planning by enabling more realistic synthetic populations that capture both demographic structure and behavioral patterns. The simultaneous multi-source integration approach represents a methodological shift from sequential processing, potentially influencing how researchers approach data fusion in other domains requiring synthetic data generation. The demonstrated improvements in diversity and feasibility metrics suggest broader applicability to fields requiring representative synthetic populations, including epidemiological modeling, economic simulation, and infrastructure planning. As synthetic data generation becomes increasingly essential for AI training (addressing insufficient data volume and quality challenges), this WGAN-based approach with regularization techniques may establish new standards for balancing realism, diversity, and computational efficiency in population synthesis.