Multi-Task Time Series Forecasting Based on Graph Neural Networks.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Accurate time series forecasting is crucial for healthcare, transportation, and finance.
• Complex, dynamic dependencies exist within time series data due to temporal variations and periodicity.
• Capturing these dependencies across historical and related tasks is challenging for predictive models.

Purpose of the Study:

• To propose a novel cross-timestep feature-sharing multi-task time series forecasting model.
• To address the challenge of capturing both global and local dynamic dependencies in time series data.
• To enhance the generalization performance of time series forecasting models.

Main Methods:

• Utilized a self-attention mechanism to capture global dynamic dependencies within each task.
• Employed an adaptive sparse graph structure to model local dynamic dependencies and inter-task correlations.
• Implemented a graph attention mechanism for cross-timestep feature sharing between related tasks.

Main Results:

• The proposed model effectively captures global and local dynamic dependencies.
• Cross-timestep feature sharing enhances the learning of strongly correlated shared features.
• Experimental results show the model is significantly competitive against baseline methods.

Conclusions:

• The developed model successfully captures intricate dependencies in time series data.
• Feature sharing across tasks and time steps improves model generalization.
• The approach offers a competitive solution for accurate multi-task time series forecasting.