Abstract
Time series clustering finds wide application but is often limited by data quality and the inherent limitations of existing methods. Compared to high-dimensional structured data like images, the low-dimensional features of time series contain less information, and endogenous noise can easily obscure important patterns. When dealing with massive time series data, existing clustering methods often focus on mining associations between sequences. However, ideal clustering results are difficult to achieve by relying solely on pairwise association analysis in the presence of noise and information scarcity. To address these issues, we propose a framework called Patch Graph Mamba (PG-Mamba). For the first time, the spatio-temporal patterns of a single sequence are explored by dividing the time series into multiple patches and constructing a spatio-temporal graph (STG). In this graph, these patches serve as nodes, connected by both spatial and temporal edges. By leveraging Mamba-driven long-range dependency learning and a decoupled spatio-temporal graph attention mechanism, our framework simultaneously captures temporal dynamics and spatial relationships and, thus, enabling the effective extraction of key information from time series. Furthermore, a spatio-temporal adjacency matrix reconstruction loss is introduced to mitigate feature space perturbations induced by the clustering loss. Experimental results demonstrate that PG-Mamba outperforms state-of-the-art methods, offering new insights into time series clustering tasks. Across the 33 datasets of the UCR time series archive, PG-Mamba achieved the highest average rank of 3.606 and secured the most first-place rankings (13).