Graph-patchformer: Patch interaction transformer with adaptive graph learning for multivariate time series...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Traditional multivariate time series (MTS) forecasting methods often overlook structural information and inter-series dependencies.
• Existing deep learning approaches may fail to capture local dynamic dependencies across series effectively.
• Multi-scale representation learning methods require additional fusion modules for final outputs.

Purpose of the Study:

• To propose a novel deep learning framework, Graph-Patchformer, for enhanced MTS forecasting.
• To address limitations in capturing structural information and inter-series local dynamic dependencies.
• To improve the accuracy and efficiency of MTS forecasting models.

Main Methods:

• Graph-Patchformer utilizes structural encodings to represent inter-series relationships and temporal variations.
• Patch Interaction Blocks with multi-head self-attention and adaptive graph learning capture dependencies.
• The framework enables interactions between patches within and across different time series.

Main Results:

• Graph-Patchformer demonstrates superior forecasting performance compared to state-of-the-art methods.
• Significant improvements were observed across various real-world benchmark datasets.
• The model effectively captures both intra-series and inter-series local dynamic dependencies.

Conclusions:

• Graph-Patchformer offers a novel and effective approach to multivariate time series forecasting.
• The framework's ability to model complex dependencies leads to state-of-the-art performance.
• This work contributes a powerful new tool for intelligent digitalization and development.