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

arXiv:2610.13730 · cs.LG · Submitted 2026-10-09 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "SeoulMMOD: A Large-Scale Multimodal Origin-Destination Flow Benchmark".

Jane: The gist SeoulMMOD introduces a large-scale public benchmark for citywide multimodal mobility in Seoul, providing three years of hourly origin-destination flows across six travel modes and two spatial levels,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, looking at SeoulMMOD: A Large-Scale Multimodal Origin–Destination Flow Benchmark, the authors are Taeyoung Yu, Seonbin Jo, Jiwon Kim, and Junyoung Byun #pg1. They introduce SEOULMMOD as a large-scale public benchmark for citywide multimodal mobility in Seoul #pg3.

Jane: It provides three years of hourly OD flows estimating total mobility for six urban travel modes across twenty-five districts and four hundred twenty-six sub-districts, together with travel-time, travel-distance, calendar, rainfall, and point-of-interest signals #pg3 <ref:2610.13730#pg1,urban travel modes across 25 districts and 426 sub-districts, together with>. That’s a lot of data for one benchmark.

Lu: The core contribution is releasing this data to benchmark fourteen forecasting baselines under a common protocol #pg3. They analyze the effects of spatial scaling, joint versus mode-separate training, and cross-year generalization #pg3.

Meng: They show that the OD flow follows a heavy-tailed distribution, which is consistent with observed human mobility patterns in cities #pg6. That means you’re dealing with some pretty extreme spikes in travel demand.

Tom: The spatial pattern is uneven, and the mode-specific network structure shows that modes differ in their spatial community structure #pg6. For instance, car trips show long-range connections spanning the city, which leads to low modularity and few large clusters with Q equals zero point five eight four and seven clusters #pg6.

Jane: That’s interesting because walk trips show short-range connections that form many localized clusters with Q equals zero point eight five two and fourteen clusters #pg6. It shows how different travel choices shape the city's structure depending on the mode you look at.

Lu: They also note that naive joint multi-mode training does not consistently improve accuracy #pg6. This suggests that simply feeding everything together isn't always the right approach for this kind of data #pg6.

Meng: And when you try to scale up, they found that several spatio-temporal models that work at a district level do not scale to sub-district OD forecasting #pg6. That’s a big warning for production systems; scaling the prediction target by two orders of magnitude means your existing infrastructure will likely break if you just try to add more spatial detail without changing how you train the model #pg10.

Tom: So, they are highlighting that we need to look beyond just fitting models on district level and figure out how to handle those finer spatial details consistently across all modes #pg6.

Jane: This paper gives us the framework and the data needed to start developing models that address these requirements for fine spatial resolution and cross-mode information #pg6.

Conclusion: Tom: So, wrapping up on SeoulMMOD, the large-scale multimodal origin-destination flow benchmark from Taeyoung Yu and his team #pg1. This paper provides three years of hourly mobility data across six modes in Seoul which is a massive dataset for testing #pg1.

Jane: It gives us three years of hourly mobility data across six modes in Seoul, which is a massive dataset for testing the limits of current forecasting systems #pg3. The core idea here is to create a reproducible testbed for forecasting citywide, multi-mode demand over multiple years #pg6.

Lu: The authors show that cross-year transfer can be tested by retraining five architectures on all of two thousand twenty-four and evaluating on all of two thousand twenty-five #pg10 <ref:2610.13730#pg2>. All five models have higher mean absolute error in cross-year evaluation, with iTransformer showing the smallest increase #pg10.

Meng: They also found that the seasonal ordering is consistent across architectures, with spring being the hardest and summer the easiest for prediction #pg10. That's a pattern we can use when setting expectations for future model performance.

Tom: The study points to STAEformer as performing best overall in terms of mean absolute error across all three prediction horizons #pg8. It reports the lowest overall MAE at all three horizons #pg8.

Jane: And they found that context signals, like holiday calendars or rainfall, actually help lower the error when added to certain models like STID over a twenty-four hour window #pg9. This is a concrete finding about how external information can help refine the forecast.

Lu: They also noted that training one model per mode outperforms joint training in average MAE for all four architectures #pg10. This suggests that future multi-mode models need mechanisms for mode-specific patterns and cross-mode signal, instead of just expanding the output dimension #pg10.

Meng: And they pointed to a limitation: forecasting OD demand at the sub-district level expands the prediction target by two orders of magnitude, from six hundred twenty-five to one hundred eighty-one thousand four hundred seventy-six OD pairs #pg10. Models that explicitly capture spatial interaction across OD pairs exceed the available memory when tested at this scale #pg10.

Tom: So, it’s a huge leap in complexity for any forecasting system trying to stay efficient and accurate while running in real time #pg3.

Jane: The paper is a reproducible testbed for developing and evaluating models that can actually handle this kind of city-scale problem #pg6. It shows the challenges we have right now with city-wide mobility data #pg6.

Taeyoung Yu, Seonbin Jo, Jiwon Kim, Junyoung Byun

University of Queensland · POSTECH · Chung-Ang University

cs.LG

Submitted: 2026-10-09

Updated: 2026-10-09

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: The gist SeoulMMOD introduces a large-scale public benchmark for citywide multimodal mobility in Seoul, providing three years of hourly origin-destination flows across six travel modes and two

Key concepts

SEOULMMOD: This is a massive public dataset tracking hourly trips within Seoul for three years (2023-2025). It covers 6 travel modes across 426 sub-districts, including travel times and weather data. It serves as a reproducible testbed to evaluate how well AI models can predict city-wide mobility.
Spatial Demand Heterogeneity: Demand for travel is not uniform across Seoul's areas. Some neighborhoods have high inflow during the morning rush hour, while others are destination-heavy. This uneven spatial pattern means models must account for different local travel behaviors when predicting flows.
Mode-Specific Network Structure: Different modes of transport—like cars versus walking—connect Seoul's zones in distinct ways. Car trips often create long-range connections, while walk trips form many short-range clusters. This structural diversity requires models to understand how each travel mode interacts with the city's geography.

Terminology

Benchmark — 1× in this paper · explained in 2 episodes

Summary

The gist SeoulMMOD introduces a large-scale public benchmark for citywide multimodal mobility in Seoul, providing three years of hourly origin-destination flows across six travel modes and two spatial levels, which establishes a reproducible testbed for multi-mode, multi-year OD forecasting at the city scale<ref:2610.13730#pg2>.

Dataset Description

SeoulMMOD is a large-scale public benchmark for citywide multimodal mobility in Seoul<ref:2610.13730#pg3>. It provides three years of hourly OD flows estimating total mobility for 6 urban travel modes across 25 districts and 426 sub-districts, together with travel-time, travel-distance, calendar, rainfall, and point-of-interest (POI) signals<ref:2610.13730#pg3>. The records are derived from publicly available “Seoul Metropolitan Area Living Mobility Data,” jointly produced by the Seoul Metropolitan Government and KT (a major South Korean mobile carrier)<ref:2610.13730#pg5>. To construct SEOULMMOD, the researchers filtered the original source records to OD pairs whose origin and destination both lie within Seoul<ref:2610.13730#pg5>. The final dataset spans 2023-01-01 through 2025-12-31, covering 1,096 days and 26,304 hourly time-steps at two spatial levels, 25 districts and 426 sub-districts within Seoul<ref:2610.13730#pg5>.

Dataset Characteristics

SEOULMMOD shows uneven spatial demand, recurrent temporal cycles, and mode-specific network structure<ref:2610.13730#pg6>. The OD flow follows a heavy-tailed distribution, consistent with observed human mobility patterns in cities<ref:2610.13730#pg6>. The spatial pattern is uneven, with net inflow separating destination-heavy sub-districts from origin-heavy sub-districts during morning rush hour<ref:2610.13730#pg6>. Multi-mode structural diversity shows that modes differ in spatial community structure, indicating that each mode connects Seoul’s zones in a different way<ref:2610.13730#pg6>. For instance, car trips show long-range connections spanning the city, which leads to low modularity and few large clusters (Q=0.584, 7 clusters), whereas walk trips show short-range connections that form many localised clusters (Q=0.852, 14 clusters)<ref:2610.13730#pg6>.

Benchmark Evaluation

The study benchmarks 14 forecasting baselines under a common protocol and analyzes the effects of spatial scaling, joint versus mode-separate training, and cross-year generalization<ref:2610.13730#pg3>. The evaluation metrics reported are Mean absolute error (MAE), Root mean squared error (RMSE), and Weighted mean absolute percentage error (wMAPE)<ref:2610.13730#pg7>.

Model Performance Findings

The experiments reveal several key challenges for current OD forecasting models in city-scale multi-mode settings<ref:2610.13730#pg6>.

several spatio-temporal models that work at district level do not scale to sub-district OD forecasting

naive joint multi-mode training does not consistently improve accuracy

The performance comparison across different scales and settings showed that STAEformer, which applies attention over the flattened OD-pair representation, reports the lowest overall MAE at all three horizons<ref:2610.13730#pg8>. Across the three horizons, learned baselines separate into patterns where STAEformer and STID are nearly horizon-invariant<ref:2610.13730#pg8>. At the sub-district level, ODMixer gives the strongest short-horizon results at 6h and 12h, while STID is strongest at 24h<ref:2610.13730#pg9>.

Contextual Signal Utility

The study evaluates how three contextual signals—holiday calendar, hourly rainfall, and per-district POI counts—change forecasting accuracy<ref:2610.13730#pg9>. Each of the three signals lowers 24h MAE by 0.49 to 1.39 when added to STID<ref:2610.13730#pg9>. However, these effects differ at shorter horizons: Holiday embedding raises MAE at h=6 and h=12 (+0.40, +0.48 MAE), while Rainfall reverses sign with horizon, going from +1.34 MAE at h=6 to −0.70 at h=12<ref:2610.13730#pg9>.

Scalability and Training Trade-offs

Forecasting OD demand at the sub-district level expands the prediction target by two orders of magnitude, from 625 to 181,476 OD pairs<ref:2610.13730#pg10>. Models that explicitly capture spatial interaction across OD pairs exceed the available memory when tested at this scale<ref:2610.13730#pg10>. Furthermore, training one model per mode outperforms joint training in average MAE for all four architectures<ref:2610.13730#pg10>. This suggests that future multi-mode models need mechanisms for mode-specific patterns and cross-mode signal, instead of just expanding the output dimension<ref:2610.13730#pg10>.

Generalization Analysis

Cross-year transfer is tested by retraining five architectures on all of 2024 and evaluating on all of 2025<ref:2610.13730#pg10>. All five models have higher MAE in cross-year evaluation, with iTransformer showing the smallest increase<ref:2610.13730#pg10>. The seasonal ordering is consistent across architectures, with spring being the hardest and summer the easiest<ref:2610.13730#pg10>.

Conclusion

SEOULMMOD highlights two central challenges for current OD forecasting models in city-scale multi-mode settings<ref:2610.13730#pg6>. Future model design should scale to all origin-destination pairs at fine spatial resolution and use cross-mode information more directly<ref:2610.13730#pg6>. SEOULMMOD provides a reproducible testbed for developing and evaluating such models<ref:2610.13730#pg6>.

Reproducibility

The project repository releases the dataset builders, model configurations, BasicTS-compatible adapters, and training and evaluation scripts used in the benchmark<ref:2610.13730#pg10>. The release includes instructions, preprocessing scripts, model configurations, training and evaluation commands<ref:2610.13730#pg10>. The dataset release provides the OD flow parquet files and metadata needed to build the benchmark tensors used for Table 2<ref:2610.13730#pg10>.

Data Provenance

SEOULMMOD is built from the Seoul Metropolitan Government’s “Seoul Metropolitan Area Living Mobility Data,” jointly produced with KT (formerly Korea Telecom)<ref:2610.13730#pg5>. The processed SEOULMMOD release is distributed under CC BY-NC 4.0, with attribution to the original Seoul Open Data Plaza source<ref:2610.13730#pg18>. The public dataset will be hosted on Kaggle<ref:2610.13730#pg18>. The underlying Seoul data are distributed under the Korea Open Government License (KOGL) Type 1<ref:2610.13730#pg5>.

Limitations

SEOULMMOD is a single-city benchmark<ref:2610.13730#pg6>. Strong performance here does not by itself establish transfer to other cities, operators, or zoning systems<ref:2610.13730#pg6>. The OD flows are upstream mobile-network estimates rather than direct counts of all trips<ref:2610.13730#pg6>. Although the source applies population scaling, estimation error may remain for visitors, phone-use patterns, or under-represented groups<ref:2610.13730#pg5>.

Experimental Setting

The paper specifies the data splits, horizons, baselines, metrics, and training protocol<ref:2610.13730#pg10>. Full model configurations and hyperparameters are provided with the released code<ref:2610.13730#pg10>.

Improvements for AI systems

A new system can perform citywide multimodal OD forecasting by integrating six transport modes (metro bus, local bus, subway, walk, car, other) across 25 districts and 426 sub-districts for three years at an hourly resolution. This system will utilize the SEOULMMOD benchmark to predict future trip volumes based on aligned travel-time and travel-distance signals.
The improved AI system can perform fine-grained spatial forecasting by moving beyond district levels to sub-district OD forecasting, addressing the finding that several spatio-temporal models that work at district level do not scale to sub-district OD forecasting.
The system will incorporate contextual signal integration by leveraging holiday calendar, hourly rainfall, and point-of-interest (POI) signals to adjust predictions; specifically, the analysis shows that these signals can lower 24h MAE by up to 1.39 when added to STID.
The system will be designed with a capacity for cross-year generalization, allowing models trained on 2024 data to be evaluated on the full 2025 calendar, testing robustness against pronounced month-to-month variability.
The architecture should prioritize capturing mode-specific structural diversity, as demonstrated by the observation that the modes differ in spatial community structure, which suggests using network representations like disparity filters and Louvain community detection to model each mode's unique connectivity.
The system should adopt a training strategy that explicitly models cross-mode interactions, as the paper notes that naive joint multi-mode training does not consistently improve accuracy, implying the need for more sophisticated mechanisms than just expanding the output dimension.
The system can be configured to utilize a model-specific approach where training one model per mode outperforms joint training in average MAE for all four architectures, suggesting an architecture that treats modes as interacting variates rather than simply expanding channels.

Abstract

Origin-destination (OD) flow forecasting supports urban mobility decisions by predicting the volume and patterns of human movement across a city. Most existing public OD benchmarks provide only a partial view of citywide mobility, covering a single transport service or at most two modes. With partial mode coverage, observed demand changes are difficult to separate into shifts across transport modes and changes in total travel. We introduce SeoulMMOD, a large-scale public benchmark for citywide multimodal mobility in Seoul. SeoulMMOD provides three years of hourly OD flows estimating total mobility for 6 urban travel modes across 25 districts and 426 sub-districts, together with travel-time, travel-distance, calendar, rainfall, and point-of-interest (POI) signals. Using SeoulMMOD, we benchmark 14 forecasting baselines under a common protocol and analyse the effects of spatial scaling, joint versus mode-separate training, and cross-year generalization. Results show that several spatio-temporal models that work at district level do not scale to sub-district OD forecasting, and that naive joint multi-mode training does not consistently improve accuracy. Together, the dataset and benchmark establish a reproducible testbed for multi-mode, multi-year OD forecasting at the city scale.