Causal discovery and inference: concepts and recent methodological advances.

Area of Science:

• Artificial Intelligence
• Causal Inference
• Machine Learning

Background:

• Automated causal inference and discovery are crucial for understanding complex systems.
• Existing methods often struggle with time series data and confounding factors.
• This paper provides a comprehensive review of current approaches and challenges.

Approach:

• Reviews fundamental concepts: manipulations, causal models, and structural equation models.
• Presents constraint-based causal discovery using conditional independence relationships.
• Focuses on identifiability in structural equation models for causal direction determination.

Key Points:

• Discusses assumptions and validity of constraint-based causal discovery.
• Demonstrates identifiability of causal direction in two-variable systems via error term independence and structural constraints.
• Highlights advances in causal discovery from time series, including handling subsampling and confounding.

Conclusions:

• Automated causal discovery offers powerful tools for scientific understanding.
• Identifiability of causal structures is achievable under specific conditions.
• Open questions remain in causal discovery, particularly for complex time series data.