Linear Fuzzy Information-Granule-Based Fuzzy C-Means Algorithm for Clustering Time Series.

Area of Science:

• Data Mining and Machine Learning
• Time Series Analysis
• Fuzzy Logic Systems

Background:

• Traditional time series clustering methods often struggle with capturing trend similarities effectively.
• Existing fuzzy C-means (FCM) algorithms have limitations in prototype adaptability during clustering.
• The need for abstract-level time series analysis necessitates granulation techniques.

Purpose of the Study:

• To design a trend-oriented-granulation-based fuzzy C-means (FCM) algorithm for abstract-level time series clustering.
• To develop a novel distance metric for comparing fuzzy information granules.
• To enable adaptable prototype lengths in the FCM clustering process.

Main Methods:

• Employs l1 trend filtering for time series segmentation and a segment merging algorithm for optimization.
• Constructs Linear Fuzzy Information Granules (LFIGs) to represent linear trends within segments.
• Utilizes a modified Dynamic Time Warping (DTW) algorithm with a novel distance measure for LFIGs.
• Introduces a granule splitting and merging algorithm for iterative prototype updates.

Main Results:

• The proposed algorithm successfully clusters time series at a granular, trend-oriented level.
• The novel distance metric effectively captures trend similarity between granular time series.
• Adaptable prototype lengths in the FCM algorithm overcome limitations of existing approaches.
• Experimental results show superior performance compared to existing time series clustering methods.

Conclusions:

• The developed LFIG-based FCM algorithm provides an effective method for trend-aware time series clustering.
• The algorithm's ability to adapt prototype lengths enhances its flexibility and performance.
• This approach offers a robust solution for analyzing diverse time series patterns and trends.