Causal Discovery for Explainable AI: A Dual-Encoding Approach

Causal relationships matter for AI explainability: The paper demonstrates that understanding why features influence model decisions—not just that they do—requires identifying causal relationships between variables, which is essential for responsible AI systems that need to justify their decisions to users and stakeholders.

Categorical data poses a technical barrier to trustworthy AI: Traditional methods for discovering causal relationships struggle with categorical variables (like gender, passenger class, or survival status) due to numerical instability, meaning many real-world datasets used in high-stakes decisions may not be properly analyzed for their causal structure.

Dual-encoding approach improves reliability: The proposed method runs causal discovery algorithms twice using different numerical representations of the same data, then combines results through majority voting—this redundancy helps ensure more stable and trustworthy identification of causal relationships compared to single-method approaches.

Results validated against established explainability methods: When tested on the Titanic dataset, the causal structures identified by this approach aligned with results from established explainable AI methods, suggesting the technique produces interpretable and consistent results that practitioners can have confidence in.

Practical path toward more transparent AI systems: By making causal discovery more reliable for real-world datasets with categorical variables, this work supports the development of AI systems that can provide clearer, more trustworthy explanations of their decisions to non-technical stakeholders.