Gaussian Adaptive Patching Powered Fully-Connected Spatial-Temporal Graph for Multivariate Time-Series Data.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Multivariate Time-Series (MTS) data is vital across many fields.
• MTS data possesses inherent spatial-temporal (ST) dependencies.
• Existing methods often fail to capture correlations between different sensors at different times (DEDT).

Purpose of the Study:

• To develop advanced Graph Neural Networks (GNNs) for comprehensive ST dependency modeling in MTS data.
• To address the limitations of existing GNNs in capturing DEDT correlations.
• To enhance MTS representation learning through improved spatial-temporal graph construction and convolution.

Main Methods:

• Proposed Fully-Connected Spatial-Temporal Graph Neural Network (FC-STGNN) with FC graph construction and convolution.
• Introduced Gaussian Adaptive Patching (GAP) for dynamic patch learning and adaptive receptive fields in GAP-STGNN.
• Utilized moving-pooling GNN layers for effective ST dependency capture.

Main Results:

• FC-STGNN and GAP-STGNN demonstrated effective capture of comprehensive ST dependencies.
• The proposed methods achieved superior performance compared to state-of-the-art (SOTA) methods on multiple MTS datasets.
• Improved FC graph construction led to enhanced MTS representation learning.

Conclusions:

• FC-STGNN and GAP-STGNN offer significant advancements in modeling ST dependencies for MTS data.
• The adaptive patching approach in GAP-STGNN further refines local pattern capture and temporal continuity.
• These novel GNN frameworks provide a more robust solution for MTS representation learning.