Structure Is Information: Structural Identifiability Mappings for Machine Learning With Partially Observed Dynamical...

Area of Science:

• Machine Learning
• Dynamical Systems
• Time Series Analysis

Background:

• Machine learning for time series classification faces challenges due to limited training data quality and quantity.
• Dynamical models offer a way to incorporate domain knowledge but can suffer from partial observability, leading to structural unidentifiability.

Purpose of the Study:

• To improve machine learning classification performance for time series data by addressing structural unidentifiability in dynamical models.
• To investigate the impact of structural identifiability (SI) analysis on classifier generalization, particularly with sparse or irregular data.

Main Methods:

• Employed structural identifiability (SI) analysis to identify and relate parameter configurations yielding identical system outputs.
• Integrated SI analysis findings into the classifier training process for time series data.
• Evaluated classification performance on example models from the biomedical domain.

Main Results:

• The proposed method significantly enhances classifier generalization to unseen data.
• Performance improvements are most notable when training datasets are limited.
• Demonstrated the practical importance of SI analysis in machine learning applications.

Conclusions:

• Structural identifiability analysis is crucial for developing robust machine learning classifiers for time series data derived from dynamical systems.
• Addressing structural unidentifiability through SI analysis enhances model interpretability and performance, especially in data-scarce scenarios.
• Highlights the need for greater attention to SI within the machine learning community.