A Constraint Optimization Approach to Causal Discovery from Subsampled Time Series Data.

Area of Science:

• Causal inference
• Time series analysis
• Machine learning

Background:

• Subsampling time series data can distort causal structure estimation.
• Existing methods struggle with accuracy when measurement timescales differ from system timescales.

Purpose of the Study:

• To develop a robust method for estimating system timescale causal structures from sub-sampled time series data.
• To improve computational efficiency and accuracy in causal discovery from aggregated time series.

Main Methods:

• A constraint satisfaction procedure for matching system and measurement timescales.
• A novel constraint optimization approach for recovering causal structure from finite-sample data.
• Application to real-world data, robustness and scalability investigations, and solver comparisons.

Main Results:

• A constraint satisfaction procedure with orders of magnitude better computational performance than prior methods.
• The first constraint optimization algorithm for recovering system timescale causal structure from sub-sampled data.
• Demonstrated optimal recovery from statistical errors and successful application to real-world data.

Conclusions:

• The proposed methods significantly advance the non-parametric estimation of causal structures from sub-sampled time series.
• The constraint optimization approach offers optimal recovery and handles statistical uncertainties effectively.
• This work provides a foundation for understanding and addressing challenges in causal discovery from aggregated time series.