Research map / Probabilistic and scientific modeling / Dynamical systems

SeoulMMOD: A Large-Scale Multimodal Origin-Destination Flow Benchmark

NeurIPS 2026 · Probabilistic and scientific modeling › Dynamical systems · approach: Data assimilation from sparse observations

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces SeoulMMOD, a three-year hourly multi-mode (six modes) origin-destination flow benchmark for Seoul at district and sub-district levels, and benchmarks 14 forecasting baselines under a common protocol.

Problem

Existing public OD forecasting benchmarks cover only one or two transport modes, are small, and span short periods, so it is hard to separate mode shifts from total travel change and to test scaling or cross-year generalization.

Approach

A citywide, six-mode hourly OD benchmark built from KT mobile-signalling-derived Seoul mobility data, bundled with travel time/distance, calendar, rainfall, and POI signals, plus a unified evaluation of spatial scaling, joint versus per-mode training, and cross-year transfer.

Key results

• At district level, STAEformer achieves the lowest 24h MAE (21.4) among learned models, about 6.1% below Historical Average (22.8).
• At sub-district scale (181,476 OD pairs), STAEformer, AGCRN, GWNet, MTGNN, MPGCN and ODCRN run out of memory and iTransformer, SOFTS, MTSMixer, TimeMixer exceed the time budget; only STID and ODMixer complete.
• Training one model per mode beats joint six-mode training in average MAE for all four tested architectures (e.g., STID 24.63 joint vs 20.34 separate, -17.4%).

Surprising findings

• Historical Average with an hour-of-week lookup is competitive with learned models, only 6.1% behind STAEformer at 24h.
• Naive joint multi-mode training does not consistently improve accuracy; per-mode models outperform joint models for all four tested architectures.

Limitations

• Single-city benchmark; does not establish transfer to other cities, operators, or zoning systems.
• OD flows are upstream mobile-network estimates rather than direct trip counts and may carry population-scaling bias.
• Only a single carrier's subscriber base underlies the source data, with residual error for visitors and under-represented groups.

Experimental setup

Datasets: SeoulMMOD (Seoul Metropolitan Area Living Mobility Data, KT signalling-derived OD flows) Seoul rainfall observations Seoul Open Data Plaza POI release

Benchmarks: SeoulMMOD

Metrics: MAE RMSE wMAPE

Baselines: Historical Average Latest Value STAEformer STID AGCRN Graph WaveNet MTGNN iTransformer SOFTS MTSMixer TimeMixer MPGCN

Builds on: STID STAEformer BasicTS ODMixer

Closest papers by meaning

• Satellites Reveal Mobility: A Commuting Origin-destination Flow Generator for Global Cities (NeurIPS 2025 · similarity 0.88)
• BO4Mob: Bayesian Optimization Benchmarks for High-Dimensional Urban Mobility Problem (NeurIPS 2025 · similarity 0.86)
• Drawing Informative Gradients from Sources: A One-stage Transfer Learning Framework for Cross-city Spatiotemporal Forecasting (AAAI 2025 · similarity 0.85)
• Taming the Long Tail in Human Mobility Prediction (NeurIPS 2024 · similarity 0.84)
• TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation (NeurIPS 2026 · similarity 0.83)
• Diffusion Transformers as Open-World Spatiotemporal Foundation Models (NeurIPS 2025 · similarity 0.83)