GeoMAE : Masking representation learning for spatio-temporal graph forecasting with missing values.

Area of Science:

• Urban intelligence systems
• Data science
• Machine learning

Background:

• Missing data is a major challenge in urban intelligence systems, impacting traffic and energy predictions.
• Existing spatio-temporal forecasting methods often neglect dynamic spatial correlations and complex missing data patterns.

Purpose of the Study:

• To develop a robust spatio-temporal learning methodology for incomplete urban datasets.
• To enhance the accuracy of traffic forecasting and energy consumption prediction models.

Main Methods:

• Introduction of GeoMAE, a spatio-temporal representation learning model.
• Utilizes an attention-based spatio-temporal forecasting network (STAFN).
• Incorporates a self-supervised auxiliary loss inspired by Masking AutoEncoders for enhanced robustness.

Main Results:

• GeoMAE significantly outperforms existing benchmark models on real-world datasets.
• Achieved up to 13.20% relative improvement in Mean Absolute Error (MAE).
• Demonstrated superior performance even with high missing data rates (e.g., 25% missing).

Conclusions:

• GeoMAE offers a robust solution for spatio-temporal learning with missing data.
• The model enhances the generalizability and accuracy of urban intelligence applications.
• Self-supervised learning is effective in improving spatio-temporal representation learning.