PG-Mamba: An Enhanced Graph Framework for Mamba-Based Time Series Clustering.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Time series clustering is crucial but challenged by data quality and method limitations.
• Low-dimensional time series features and noise hinder pattern discovery.
• Existing methods often rely on pairwise associations, struggling with massive datasets.

Purpose of the Study:

• To introduce a novel framework, Patch Graph Mamba (PG-Mamba), for improved time series clustering.
• To address limitations of existing methods in handling noisy and information-scarce time series data.
• To explore spatio-temporal patterns within individual time series for enhanced clustering.

Main Methods:

• Dividing time series into patches to construct a spatio-temporal graph (STG).
• Utilizing Mamba for long-range dependency learning and a graph attention mechanism.
• Incorporating a spatio-temporal adjacency matrix reconstruction loss to stabilize feature space.

Main Results:

• PG-Mamba demonstrates superior performance over state-of-the-art time series clustering methods.
• Achieved the highest average rank (3.606) across 33 UCR archive datasets.
• Secured the most first-place rankings (13) in time series clustering tasks.

Conclusions:

• PG-Mamba effectively extracts key information from time series by capturing spatio-temporal dynamics.
• The framework offers a new approach to time series clustering, particularly for noisy and low-dimensional data.
• PG-Mamba provides significant advancements and new insights in the field of time series analysis.