Exploring dual-view graph structures: Contrastive learning with graph and hypergraph for multivariate time series...

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Multivariate time series classification requires capturing both temporal dynamics and inter-variable relationships.
• Existing graph-based methods often overlook high-order, non-pairwise variable dependencies.
• Graph complexity can introduce noise, hindering the identification of crucial local patterns.

Purpose of the Study:

• To propose a novel framework, Dual-View Graph-structured Contrastive Learning (DVG-CL), for enhanced multivariate time series classification.
• To model multivariate time series using both graph and hypergraph structures to capture diverse relationships.
• To address limitations in existing methods by incorporating high-order dependencies and reducing noise.

Main Methods:

• DVG-CL models time series as both a graph and a hypergraph to capture low-order pairwise and high-order non-pairwise relationships.
• A cross-view contrasting loss is employed to synergize relationships across different structural levels.
• A local-global mutual information loss is introduced to filter noise and highlight key local aggregation information.

Main Results:

• DVG-CL demonstrated superior performance compared to existing self-supervised learning baselines on 11 UEA datasets.
• Experimental results validated the effectiveness of individual components within the DVG-CL framework.
• The model successfully captured complex inter-variable dependencies, including high-order relationships.

Conclusions:

• The proposed DVG-CL framework offers a robust approach to multivariate time series classification.
• Modeling time series with both graph and hypergraph structures enhances the capture of complex dependencies.
• DVG-CL effectively mitigates noise and identifies critical local information, leading to improved classification accuracy.