Toeplitz Inverse Covariance-Based Clustering of Multivariate Time Series Data

Area of Science:

• Machine Learning
• Data Mining
• Time Series Analysis

Background:

• Subsequence clustering of multivariate time series aids in discovering temporal data patterns.
• Interpreting complex datasets as sequences of states (clusters) simplifies analysis.
• Simultaneous segmentation and clustering of time series, especially high-dimensional data, presents significant challenges.

Purpose of the Study:

• To propose a novel model-based clustering method for multivariate time series.
• To address the challenges of simultaneous segmentation and clustering in high-dimensional temporal data.
• To develop a method that facilitates the interpretation of discovered temporal patterns.

Main Methods:

• Introduced Toeplitz Inverse Covariance-based Clustering (TICC), a model-based clustering approach.
• Defined clusters using correlation networks (Markov random fields) to represent interdependencies.
• Employed alternating minimization, a variation of the Expectation-Maximization (EM) algorithm, to solve the TICC problem.

Main Results:

• Developed efficient, closed-form solutions for subproblems using dynamic programming and ADMM.
• Demonstrated TICC's effectiveness through comparisons with state-of-the-art methods on synthetic datasets.
• Showcased TICC's ability to learn interpretable clusters on a real-world automobile sensor dataset.

Conclusions:

• TICC offers a scalable and effective solution for simultaneous segmentation and clustering of multivariate time series.
• The method successfully identifies and characterizes temporal patterns, enabling better data interpretation.
• TICC holds promise for analyzing complex temporal data across various domains, including sensor data analysis.