Causal network discovery by iterative conditioning: Comparison of algorithms.

Area of Science:

• Complex systems analysis
• Network science
• Causal inference

Background:

• Estimating causal interactions in complex dynamical systems is crucial across scientific fields.
• Existing prediction improvement frameworks face the curse of dimensionality with increasing network size.
• Iterative methods assessing conditional independences offer potential solutions.

Purpose of the Study:

• To theoretically and numerically compare prominent methods for causal interaction estimation.
• To analyze algorithm assumptions, strengths, weaknesses, and properties like false positive control.
• To identify computationally efficient yet accurate approaches for complex network analysis.

Main Methods:

• Theoretical comparison of algorithms within a unified framework.
• Numerical simulations using realistic complex coupling patterns.
• Evaluation of accuracy, computational demands, and order-dependence.

Main Results:

• Theoretical analysis revealed key similarities and differences among algorithms.
• Numerical simulations showed comparable accuracy but significant differences in computational complexity (polynomial to exponential).
• Computational demands varied substantially based on network density and size.

Conclusions:

• A hybrid approach is proposed, combining competitive accuracy with enhanced computational efficiency.
• This hybrid method addresses the computational challenges in large-scale causal network inference.
• The findings guide the selection of appropriate methods for causal discovery in complex systems.