A Kernel Embedding-Based Approach for Nonstationary Causal Model Inference.

Area of Science:

• Machine Learning
• Causal Inference
• Statistics

Background:

• Most causal discovery methods assume stationary data, which is unrealistic in real-world scenarios.
• Nonstationary data, common in real-world applications, presents challenges for existing causal inference techniques.
• Existing methods often fail to account for varying data distributions across different domains.

Purpose of the Study:

• To propose a novel kernel embedding-based approach, ENCI, for causal model inference under nonstationarity.
• To develop a method capable of discovering causal relationships from data collected across multiple domains with differing distributions.
• To extend the approach for causal graph discovery involving multiple variables.

Main Methods:

• ENable Nonstationary Causal Inference (ENCI) transforms cause-effect relationships into a linear model using kernel embeddings of distributions.
• The method exploits causal asymmetry in the transformed linear model to estimate causal direction.
• ENCI is extended to causal graph discovery by modeling relations as a linear non-Gaussian acyclic model.

Main Results:

• ENCI demonstrates identifiability of cause-effect pairs and causal graphs under mild conditions by exploiting nonstationarity.
• The proposed method shows superior performance compared to major existing methods on both synthetic and real-world datasets.
• Kernel embedding of distributions effectively captures nonstationarity for causal inference.

Conclusions:

• ENCI provides an effective framework for causal discovery in nonstationary environments.
• The method's ability to handle varying data distributions across domains enhances its practical applicability.
• Exploiting nonstationarity is crucial for accurate causal inference in complex, real-world systems.