ST-CRMF: Compensated Residual Matrix Factorization with Spatial-Temporal Regularization for Graph-Based Time Series...

Area of Science:

• * Transportation Science
• * Data Science
• * Machine Learning

Background:

• * Accurate traffic time series forecasting remains a significant challenge despite extensive research.
• * Existing models often struggle with the non-linear dynamics and spatial-temporal dependencies inherent in traffic data.
• * The issue of missing traffic data is frequently overlooked in current prediction models.

Purpose of the Study:

• * To propose a novel graph-based model for traffic time series forecasting that addresses non-linearity and spatial-temporal correlations.
• * To develop a model capable of handling missing traffic data while performing accurate predictions.
• * To improve the robustness and accuracy of traffic forecasting, particularly in short-to-long term horizons.

Main Methods:

• * Introduction of the Compensated Residual Matrix Factorization with Spatial-Temporal (ST-CRMF) regularization model.
• * Utilization of a bi-directional residual structure for compensatory modeling of spatial-temporal correlations.
• * Synchronized iterative updates for matrix factorization (MF) modeling and residual learning to mitigate error propagation.

Main Results:

• * The ST-CRMF model effectively captures comprehensive spatial-temporal dependencies in traffic data.
• * Synchronized updates alleviate the error propagation issue common in rolling forecasting.
• * The model demonstrates superior performance compared to state-of-the-art methods on Seattle-Loop and METR-LA datasets for 5- to 60-minute forecasts.
• * The ST-CRMF model successfully repairs missing traffic data while maintaining forecasting accuracy.

Conclusions:

• * The proposed ST-CRMF model offers a significant advancement in traffic time series forecasting.
• * The model's ability to handle non-linearity, spatial-temporal dependencies, and missing data makes it highly effective.
• * ST-CRMF provides a robust and accurate solution for both short-term and long-term traffic prediction tasks.