Bivariate Causal Discovery for Categorical Data via Classification with Optimal Label Permutation.

Area of Science:

• Causal inference
• Machine learning
• Categorical data analysis

Background:

• Causal discovery is well-established for quantitative data but remains challenging for categorical data.
• Existing methods for categorical causal discovery often lack identifiability or parsimony.

Purpose of the Study:

• To propose a novel, identifiable, and parsimonious causal model specifically designed for categorical data.
• To develop a straightforward algorithm for learning causal directions in categorical datasets.
• To evaluate the performance of the proposed model against current state-of-the-art methods.

Main Methods:

• Development of a new classification model: Classification with Optimal Label Permutation (COLP).
• Leveraging COLP's parsimonious nature to ensure a provably identifiable causal model.
• Employing a likelihood comparison between causal and anti-causal models for learning causal direction.

Main Results:

• The proposed COLP-based causal model demonstrates favorable performance in experiments.
• The method effectively learns causal directions for categorical data.
• Empirical validation using both synthetic and real-world datasets confirms the model's efficacy.

Conclusions:

• The COLP-based causal discovery model offers a significant advancement for categorical data.
• The developed algorithm provides a practical approach to identifying causal relationships.
• An R package (COLP) is released, including the algorithm and a benchmark dataset for future research.