Namwoo Kim

Urban region embeddings don't need sprawling multimodal datasets to hit state-of-the-art—modeling dynamic topological persistence and temporal flow synergy within mobility data alone is enough to outperform multi-source baselines.

Multi-scale temporal hierarchies don't reliably emerge on their own in Euclidean latent spaces, but forcing representations into a Poincaré ball via radial and angular geometric constraints unlocks state-of-the-art long-term forecasting.