Analysis of switching dynamics with competing support vector machines

Area of Science:

• Machine Learning
• Time Series Analysis
• Signal Processing

Background:

• Nonstationary time series analysis often requires segmentation of dynamic changes.
• Previous methods utilized annealed competing neural networks for time series segmentation.
• Limitations exist in adapting segmentation parameters for evolving data patterns.

Purpose of the Study:

• To develop an unsupervised framework for segmenting switching dynamics in nonstationary time series.
• To adapt the architecture of Pawelzik et al. using support vector machines (SVMs).
• To introduce a novel formulation for support vector regression and an adaptive annealing parameter adjustment method.

Main Methods:

• Utilized support vector machines (SVMs) for unsupervised time series segmentation.
• Proposed a new formulation of support vector regression.
• Implemented an expectation-maximization (EM) algorithm for adaptive annealing parameter adjustment.
• Applied the framework to segment nonstationary time series data.

Main Results:

• The proposed SVM-based framework demonstrated effective unsupervised segmentation of switching dynamics.
• The adaptive adjustment of the annealing parameter improved segmentation performance.
• Results indicate the framework is a promising alternative for nonstationary time series analysis.

Conclusions:

• The developed framework offers a robust method for unsupervised segmentation of switching dynamics.
• Support vector machines provide a viable alternative to neural networks for this task.
• Further research can explore extensions and applications of this adaptive segmentation approach.