Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Causal Discovery for Explainable AI: A Dual-Encoding Approach

Henry Salgado, Meagan R. Kendall, Martine Ceberio

TL;DR

This work tackles causal discovery for explainable AI in mixed-type data by introducing a dual-encoding approach that mitigates numerical instability in conditional independence testing. By running the FCI algorithm on both Drop-first and Drop-last encodings and merging results through majority voting, the method yields unified causal graphs that align with domain knowledge and SHAP-based feature importance on the Titanic dataset. The approach provides directional and mediating insights beyond standard feature-importance measures, enabling more robust global explanations and paving the way for potential instance-level reasoning and counterfactual analyses. Its practical impact lies in offering a stable, interpretable causal framework for complex datasets with mixed data types, with demonstrated coherence between learned structures and established explanations.

Abstract

Understanding causal relationships among features is fundamental for explaining machine learning model decisions. However, traditional causal discovery methods face challenges with categorical variables due to numerical instability in conditional independence testing. We propose a dual-encoding causal discovery approach that addresses these limitations by running constraint-based algorithms with complementary encoding strategies and merging results through majority voting. Applied to the Titanic dataset, our method identifies causal structures that align with established explainable methods.

Causal Discovery for Explainable AI: A Dual-Encoding Approach

TL;DR

This work tackles causal discovery for explainable AI in mixed-type data by introducing a dual-encoding approach that mitigates numerical instability in conditional independence testing. By running the FCI algorithm on both Drop-first and Drop-last encodings and merging results through majority voting, the method yields unified causal graphs that align with domain knowledge and SHAP-based feature importance on the Titanic dataset. The approach provides directional and mediating insights beyond standard feature-importance measures, enabling more robust global explanations and paving the way for potential instance-level reasoning and counterfactual analyses. Its practical impact lies in offering a stable, interpretable causal framework for complex datasets with mixed data types, with demonstrated coherence between learned structures and established explanations.

Abstract

Understanding causal relationships among features is fundamental for explaining machine learning model decisions. However, traditional causal discovery methods face challenges with categorical variables due to numerical instability in conditional independence testing. We propose a dual-encoding causal discovery approach that addresses these limitations by running constraint-based algorithms with complementary encoding strategies and merging results through majority voting. Applied to the Titanic dataset, our method identifies causal structures that align with established explainable methods.
Paper Structure (15 sections, 4 figures)

This paper contains 15 sections, 4 figures.

Table of Contents

  1. Introduction
  2. Background
  3. Explanation Methods
  4. Causal Discovery Methods
  5. Methods
  6. Dual-Encoding Strategy
  7. Data Preparation
  8. Graph Merging and Correlation Weighting
  9. Experiments and Results
  10. Datasets
  11. Titanic Results
  12. Validation Against Baseline Methods
  13. Discussion
  14. Limitations
  15. Conclusion and Future Work

Figures (4)

  • Figure 1: Causal graphs for the Titanic dataset using drop-first (left) and drop-last (right) encoding strategies. Structural consistency across encodings demonstrates robustness to categorical reference choice.
  • Figure 2: Unified causal graph for the Titanic dataset obtained by merging encoding-specific graphs via majority voting and correlation weighting.
  • Figure 3: Pruned decision tree trained on the Titanic dataset. The root split on Sex, followed by Pclass and Age, aligns with the causal relationships identified by FCI.
  • Figure 4: SHAP summary plot illustrating global feature importance for the Titanic dataset. The most influential features align with those identified in the unified causal graph.