Self-Supervised Time Series Clustering With Model-Based Dynamics.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Time series clustering is a crucial unsupervised learning task, especially when category labels are absent.
• Existing methods often fail to adequately address temporal dynamics or integrate feature extraction with clustering.
• This limitation hinders the effective analysis of complex time series data.

Purpose of the Study:

• To propose a novel framework, the self-supervised time series clustering network (STCN), for simultaneous feature extraction and clustering.
• To enhance the performance of time series clustering by considering the interaction between feature representation and cluster assignment.
• To improve the unsupervised analysis of time series data.

Main Methods:

• STCN employs a recurrent neural network (RNN) for one-step time series prediction to capture temporal dynamics and local structures.
• Model-based dynamic features from the RNN's output layer are fed into a self-supervised clustering module.
• Spectral analysis is utilized to align predicted cluster labels with pseudo-labels, bridging feature extraction and clustering.

Main Results:

• Experiments on extensive datasets demonstrate that STCN achieves state-of-the-art performance in time series clustering.
• The proposed model effectively optimizes feature extraction and clustering simultaneously.
• Visualization analyses confirm the model's ability to uncover meaningful clusters in time series data.

Conclusions:

• STCN offers a significant advancement in unsupervised time series clustering by integrating feature learning and clustering.
• The framework's ability to capture temporal dynamics and local structures leads to superior performance.
• STCN provides an effective solution for analyzing unlabeled time series data across various applications.