Adaptive Clustering and Feature Selection for Categorical Time Series Using Interpretable Frequency-Domain Features.

Area of Science:

• Data Science
• Time Series Analysis
• Signal Processing

Background:

• Categorical time series analysis often faces challenges in pattern identification and feature extraction.
• Interpretable frequency-domain features offer a promising avenue for characterizing cyclical patterns.

Purpose of the Study:

• To develop a novel approach for clustering and feature selection in categorical time series.
• To introduce a distance measure based on spectral characteristics for time series comparison.
• To apply these methods for identifying sleep disruption patterns in patients.

Main Methods:

• A novel distance measure utilizing spectral envelope and optimal scalings for categorical time series.
• Development of partitional clustering algorithms incorporating simultaneous feature selection.
• Adaptive procedures for fuzzy membership assignment in clustering.
• Investigation of clustering consistency and simulation studies for accuracy assessment.

Main Results:

• Accurate clustering of categorical time series using the proposed spectral-based distance measure.
• Effective simultaneous feature selection to identify distinguishing characteristics of clusters.
• Demonstrated clustering accuracy across various group structures in simulation studies.
• Successful application to sleep stage time series for identifying sleep disruption patterns.

Conclusions:

• The proposed frequency-domain approach provides an effective and interpretable method for categorical time series clustering and feature selection.
• This technique can identify specific oscillatory patterns associated with sleep disruption in patient data.
• The methods offer robust performance and potential for broader applications in time series analysis.