Zero-Inflated Time Series Clustering Via Ensemble Thick-Pen Transform.

Area of Science:

• Data Science
• Time Series Analysis
• Statistical Modeling

Background:

• Clustering high-dimensional time series data, especially with zero inflation, presents significant challenges.
• Existing methods often struggle with the temporal dependencies and sparsity inherent in such data.
• Effective clustering is vital for uncovering patterns in complex datasets.

Purpose of the Study:

• To develop a novel clustering method for high-dimensional zero-inflated time series data.
• To enhance the temporal resolution of time series data for improved clustering.
• To introduce a robust similarity measure and an efficient clustering algorithm tailored for this data type.

Main Methods:

• Development of an ensemble thick-pen transform (e-TPT) to improve temporal resolution.
• Definition of a modified similarity measure incorporating e-TPT for zero-inflated data.
• Proposal of an efficient iterative clustering algorithm optimized for the new similarity measure.

Main Results:

• The proposed e-TPT based clustering method demonstrates superior performance in simulation experiments.
• Effective clustering of real-world datasets, including step count data and daily COVID-19 case data, was achieved.
• The method successfully addresses the challenges of high dimensionality and zero inflation in time series.

Conclusions:

• The ensemble thick-pen transform (e-TPT) offers a powerful approach for clustering zero-inflated time series.
• The developed method provides enhanced temporal resolution crucial for accurate time series clustering.
• This technique has practical applications in analyzing health and activity-related time series data.