Multiview Unsupervised Shapelet Learning for Multivariate Time Series Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Time Series Analysis

Background:

• Multivariate time series clustering is crucial for identifying patterns in complex data.
• Existing methods often struggle to extract informative subsequences (shapelets) from multivariate time series.
• Discovering correlations and partitioning data into meaningful subsets remains a challenge.

Purpose of the Study:

• To propose a novel unsupervised method for learning salient multivariate subsequences (shapelets).
• To develop a multiview approach that leverages shapelet-transformed representations of varying lengths.
• To improve the accuracy and robustness of multivariate time series clustering.

Main Methods:

• Introduced Unsupervised Shapelet Learning with Adaptive Neighbors (USLA) for learning multivariate shapelets.
• Developed Multiview USLA (MUSLA) by treating shapelet-transformed representations from different lengths as distinct views.
• MUSLA simultaneously learns view importance and multiview neighbor graph matrices.

Main Results:

• USLA effectively learns multivariate shapelets with auto-determined variate importance.
• MUSLA integrates complementary information from shapelets of different lengths.
• Experimental results demonstrate MUSLA's superior performance over state-of-the-art algorithms on real-world datasets.

Conclusions:

• The proposed MUSLA model offers a significant advancement in multivariate time series clustering.
• Leveraging multiview representations derived from shapelets of varying lengths enhances pattern discovery.
• MUSLA provides a robust and effective solution for partitioning complex multivariate time series data.