Locally Linear Continual Learning for Time Series Based on VC-Theoretical Generalization Bounds.

Area of Science:

• Machine Learning
• Time Series Analysis
• Statistical Learning Theory

Background:

• Traditional machine learning models struggle with nonstationary data due to fixed probability distributions.
• Existing continual learning methods often lack interpretability or require significant user input.
• The need for explainable and adaptive models in dynamic environments is critical.

Purpose of the Study:

• To develop an explainable model for time series forecasting in nonstationary environments.
• To address the limitations of black-box models and current interpretable methods.
• To reconcile predictive accuracy with model transparency.

Main Methods:

• Proposed SyMPLER (Systems Modeling through Piecewise Linear Evolving Regression), a novel approach for time series forecasting.
• Utilized dynamic piecewise-linear approximations for modeling evolving data patterns.
• Applied generalization bounds from Statistical Learning Theory to automatically manage model complexity based on prediction errors.

Main Results:

• SyMPLER achieved performance comparable to black-box and existing explainable models.
• The model demonstrated a human-interpretable structure, providing insights into system dynamics.
• Automatic model adaptation based on prediction errors eliminated the need for explicit data clustering.

Conclusions:

• SyMPLER offers a transparent and adaptive solution for forecasting nonstationary time series.
• The approach successfully integrates accuracy and interpretability in machine learning.
• This method provides valuable insights into system behavior in dynamic settings.