Time-Series Clustering Based on the Characterization of Segment Typologies.

Area of Science:

• Data Mining
• Machine Learning
• Time Series Analysis

Background:

• Traditional time-series clustering methods often overlook the similarity of subsequences within time series.
• Existing approaches typically combine a distance measure with a standard clustering algorithm, limiting comparative analysis.

Purpose of the Study:

• To propose a novel two-stage time-series clustering technique that accounts for subsequence similarities.
• To improve the accuracy and effectiveness of grouping time series data.

Main Methods:

• A least-squares polynomial segmentation procedure is applied to each time series using a growing window.
• Segments are projected into a common dimensional space using model coefficients and statistical features.
• A hierarchical clustering phase groups segments, followed by a second stage to cluster the time series objects.

Main Results:

• The proposed method was evaluated on 84 datasets from the UCR Time Series Classification Archive.
• Performance was compared against three state-of-the-art time-series clustering methods.
• The methodology demonstrated very promising performance, particularly on larger datasets.

Conclusions:

• The novel two-stage clustering approach effectively captures subsequence similarities for improved time-series grouping.
• The technique offers a promising alternative to existing methods, especially for large-scale time-series data analysis.