Iterative Reservoir Computing Networks for Reconstructing Irregular Time Series

Area of Science:

• Complex Systems and Data Science
• Time Series Analysis and Dynamical Systems

Background:

• Missing data in time series is a common challenge across diverse fields like medicine and climatology, hindering data mining and analysis.
• Existing methods often focus on interpolation or task-specific adaptations, leaving a gap for generalizable irregular time series recovery.

Purpose of the Study:

• To develop an iterative learning method based on reservoir computing (RC) for systematically recovering missing data in irregular time series.
• To formulate the data recovery as a fixed-point iterative learning problem solvable with an RC network (RCN).

Main Methods:

• Developed an iterative learning procedure using an RC network (RCN) to address missing data in irregular time series.
• Formulated the problem as a fixed-point iterative learning task.
• Derived conditions for reservoir parameters to ensure convergence of the iterative procedure.

Main Results:

• Demonstrated successful systematic recovery of missing data in irregular time series when sufficient samples are available for RCN training.
• Validated the approach on chaotic Rössler and Kuramoto-Sivashinsky (KS) systems, showcasing its efficacy.
• Showcased the method's applicability by incorporating it into an irregular medical data classification task.

Conclusions:

• The proposed iterative RCN approach offers a robust and systematic solution for recovering missing data in irregular time series from dynamical systems.
• The method shows promise for practical applications, including complex system analysis and medical data processing.