Rate-Agnostic (Causal) Structure Learning.

Area of Science:

• Time series analysis
• Causal inference
• Machine learning

Background:

• Causal structure learning from time series data is complex.
• Existing methods often assume synchronized system and measurement timescales.
• Real-world data frequently involves undersampling, where measurement rates are slower than system dynamics, and the timescale mismatch is unknown.

Purpose of the Study:

• To develop novel algorithms for causal structure learning from time series data.
• To address the challenge of undersampling in causal discovery.
• To learn causal structure in a rate-agnostic manner, without assuming specific timescale relationships.

Main Methods:

• Development of three new causal structure learning algorithms.
• Algorithms are designed to be rate-agnostic, handling unknown timescale mismatches.
• Application of algorithms to simulated data to analyze undersampling effects.

Main Results:

• The developed algorithms can discover all dynamic causal graphs consistent with observed data, even under undersampling.
• Demonstrated the capability of rate-agnostic causal learning.
• Gained insights into the impact of undersampling on causal discovery through simulations.

Conclusions:

• The proposed algorithms offer a robust solution for causal structure learning from undersampled time series data.
• Rate-agnostic causal discovery is feasible and valuable in scenarios with unknown timescale differences.
• This work advances the field of causal inference in the presence of measurement limitations.