Quantifying causal effects to enhance explainability in Causal Bayesian Networks
Rafael Arone, Henrique O. Caetano, Carlos Dias Maciel · Knowledge-Based Systems · 2026
The widespread adoption of complex, black box machine learning models has created a need for explainability methods to ensure transparency, fairness, and trust. While many techniques focus on counterfactual reasoning, a gap remains for strategies that can directly quantify the influence of variables and their specific states on an outcome. This paper introduces a unified framework for enhancing explainability in Causal Bayesian Networks by leveraging causal interventions and root cause analysis. We propose three novel methods for generating non-comparative, global explanations in models with discrete and multicategorical data, assuming causal sufficiency and the model is a directed acyclic graph. The Global Causal Impact quantifies the total influence a variable has on a target. The Path-Specific Causal Impact decomposes this influence, measuring how it propagates through specific intermediate pathways. Finally, the Explainable Root Cause Analysis identifies the configuration of variable states most likely to have produced a given effect. We validate our framework on synthetic datasets with known linear and non-linear mechanisms and demonstrate its practical utility on the several benchmark datasets. The results show that our methods provide a multifaceted, intuitive, and causally grounded understanding of a model’s predictions, successfully identifying key drivers and their underlying mechanisms.