HuMob Challenge 2026

Welcome to the Human Mobility Prediction Challenge 2026

The Human Mobility Prediction Challenge, also known as the HuMob Challenge, is a competition aimed at testing state-of-the-art computational models for predicting human mobility patterns. It follows the success of HuMob Challenge 2023 (Hamburg), HuMob Challenge 2024 (Atlanta), and GISCUP 2025 (Minnesota).

Understanding, modeling, and predicting human mobility in urban and regional areas is an essential task for many domains and applications, including transportation modeling, disaster risk management, emergency response, urban planning, and regional recovery analysis.

Recent advances in large-scale mobility data and computational modeling have enabled researchers to develop increasingly complex human mobility models. However, mobility prediction methods are often trained and tested on different datasets, making it difficult to compare model performance fairly. The HuMob Challenge provides a shared benchmark task and dataset to support reproducible evaluation of human mobility models.

The Challenge: Recovery-period OD Matrix Prediction

The 2026 challenge focuses on human mobility recovery after the 2024 Noto Peninsula Earthquake in Ishikawa Prefecture, Japan. Participants will estimate how and to what extent human mobility recovered during the post-disaster recovery period.

The task is to predict origin-destination matrices during the test period using the provided training-period OD matrices and publicly accessible contextual information, such as news articles, public reports, and official recovery-related information.

2024 Noto Peninsula Earthquake

The challenge is motivated by the severe disruption and long recovery process following the 2024 Noto Peninsula Earthquake. The images below provide visual context for the disaster-affected region and the recovery challenges that shaped post-disaster mobility patterns.

Earthquake damage in the Noto Peninsula

Photos by Kazuto Ataka.

Dataset

The HuMob Challenge 2026 dataset contains aggregated origin-destination mobility flows for the Noto / Ishikawa study area. The dataset is available on Zenodo.

Dataset File

Spatial Grid

Training and Test Periods

Missing / NA Days

Due to data quality reasons, the following 16 dates were removed and replaced with NA:

20231126, 20231130, 20231201, 20231203, 20231204, 20231205, 20231214, 20240118, 20240123, 20240124, 20240202, 20240305, 20240408, 20240426, 20240529, 20240708.

Privacy and Normalization

The dataset is anonymized using k-anonymity to protect privacy. The k value cannot be disclosed for privacy reasons. The data is also normalized by a constant number for business reasons.

Rules

Participants may use the provided training data and publicly accessible external information to predict mobility flows during the test period.

Reference Information

The organizing team will provide a list of suggested public references. Participants may also use other publicly available sources, as long as they comply with the challenge rules.

• 2024 Noto Peninsula Earthquake, Wikipedia
• Noto Peninsula Earthquake Archive, Ishikawa Prefecture

Evaluation

Submissions will be evaluated by comparing predicted and observed origin-destination matrices during the test period. The evaluation will be conducted within the evaluation boundary, defined by grid longitude range x=30 to x=70 and grid latitude range y=35 to y=70.

Primary evaluation metric: Combined NRMSE

The primary evaluation metric is the Combined Normalized Root Mean Squared Error (Combined NRMSE). The mobility data consist of daily origin-destination (OD) flows on a 2 km grid. OD pairs are evaluated separately as:

• Diagonal pairs: origin and destination are the same grid cell, representing people remaining within the same cell.
• Off-diagonal pairs: origin and destination are different grid cells, representing movement between cells.

Because diagonal flows are approximately ten times larger than off-diagonal flows on average, their raw errors are not directly comparable. The metric therefore normalizes the two components before combining them.

Computation

1. Daily RMSE: For each day in the evaluation period, the root mean squared error is calculated separately across all diagonal and off-diagonal OD pairs:

   RMSE_diag(d) = sqrt(mean((prediction - observation)²) over diagonal pairs)
   
   RMSE_offdiag(d) = sqrt(mean((prediction - observation)²) over off-diagonal pairs)

2. Mean RMSE: The daily RMSE values for each component are averaged across all valid evaluation days:

   RMSE_diag = mean(RMSE_diag(d) over days)
   
   RMSE_offdiag = mean(RMSE_offdiag(d) over days)

3. Normalization: Each mean RMSE is divided by the mean of all positive observed values for that component across the full evaluation period:

   NRMSE_diag = RMSE_diag / mean_actual_diag
   
   NRMSE_offdiag = RMSE_offdiag / mean_actual_offdiag

   Here, we set mean_actual_diag = 207.6 and mean_actual_offdiag = 19.7.

4. Combined NRMSE: The final score is the simple average of the two normalized components:

   Combined NRMSE = (NRMSE_diag + NRMSE_offdiag) / 2

Rationale

Normalization ensures that errors in diagonal and off-diagonal flows contribute equally to the final score despite their substantial difference in magnitude. Without normalization, the larger diagonal flows would dominate the evaluation. The mean observed value, rather than the median, is used as the normalization factor because the diagonal-flow distribution is strongly right-skewed; median-based normalization would disproportionately inflate the diagonal NRMSE. Lower Combined NRMSE values indicate better predictive performance.

Baseline Models

Baseline results will be provided for comparison. Planned baselines include:

• Using April 2024 OD flows directly as predictions for the test period.
• Using corresponding 2023 OD flows directly as predictions for the test period.

Workshop @ ACM SIGSPATIAL

Top performers will be invited to present their work at the HuMob Challenge Workshop @ SIGSPATIAL 2026 held in Riverside, CA, on November 3rd, 2026, and their contributions will be accepted in the Workshop Proceedings.

Organizers

Takahiro Yabe, New York University
Kota Tsubouchi, LY Corporation
Toru Shimizu, LY Corporation