Recurrence microstates for machine learning classification

Area of Science:

• Complex Systems Analysis
• Time Series Data Mining
• Machine Learning

Background:

• Recurrence quantification analysis (RQA) traditionally analyzes time series dynamics.
• Phase space recurrence is a fundamental concept in dynamical systems theory.
• Limitations exist in detecting subtle pattern changes with existing RQA methods.

Purpose of the Study:

• Introduce recurrence microstates as a generalization of phase space recurrence.
• Develop a novel feature generation tool for machine learning from time series data.
• Enhance the classification of chaotic system parameters using deep neural networks.

Main Methods:

• Obtain recurrence microstates from cross-recurrence of embedded value sequences.
• Utilize microstate occurrence probabilities as recurrence quantifiers.
• Implement a microstate multi-layer perceptron (MMLP) for parameter classification.

Main Results:

• Microstate probabilities detect subtle changes in data patterns.
• The MMLP effectively classifies parameters of chaotic systems.
• Increasing the number of microstates improves MMLP classification accuracy.

Conclusions:

• Recurrence microstates offer a sensitive and informative approach to time series analysis.
• The MMLP demonstrates strong performance in classifying chaotic system parameters.
• The method shows potential for diverse applications across different data contexts.