Causal Explanations and XAI

The evolution of explainable AI (XAI) has transitioned from simple feature importance to causal reasoning frameworks. Pearl’s causal hierarchy—observation, intervention, and counterfactuals—provides a structured approach to understanding causality. Prior work focused on correlational explanations, which are insufficient for action guidance. Structural causal models (SCMs) have become central for formal causal reasoning, enabling precise causal inference. Recent efforts integrate deep learning with causal inference, but formal definitions of causation tailored for explainability remain limited. This paper builds on these foundations, formalizing layered causal explanations—sufficient, counterfactual, and actual causation—within SCMs, aiming to improve both interpretability and fairness in AI systems.

Existing XAI methods predominantly rely on observational correlations, which often mislead decision-makers about causality. They lack formal definitions of causation, making it difficult to identify true causes and guide effective interventions. This limitation hampers fairness, especially in sensitive domains like healthcare and finance, where understanding the root causes of outcomes is critical. The core challenge is to develop a formal, scalable framework that accurately captures causality, supports action-guiding explanations, and addresses path-specific fairness, all within practical constraints.

This paper’s key innovations include: 1) formal definitions of sufficient, counterfactual, and actual causation grounded in structural causal models; 2) layered explanation framework aligning with Pearl’s hierarchy, distinguishing between different types of causal reasoning; 3) a novel formalization of path-specific fairness based on actual causation, enabling fairer decision-making. These contributions provide a rigorous, unified approach to causal explanation, surpassing heuristic or purely statistical methods, and directly supporting actionable insights.

• �� Model the target system with a deterministic structural causal model (SCM) using structural equations.
• �� Define sufficient explanations as variable settings that guarantee outcome Y under no intervention on other variables.
• �� Introduce layered notions: weak sufficiency (no intervention needed), direct sufficiency (interventions only on explanatory variables), and strong sufficiency (including safeguards against interventions).
• �� Formalize counterfactual explanations by manipulating variables and analyzing their influence on Y, considering witnesses and causal paths.
• �� Establish actual causation by minimality and counterfactual dependence, formalized through witness sets and causal diagrams.
• �� Integrate path-specific fairness by identifying causal pathways responsible for unfair influence.
• �� Validate the framework through synthetic and real datasets, comparing with baseline feature importance methods.

Using synthetic datasets with known causal structures, the model’s ability to identify true causes and guide actions was tested. Real datasets from healthcare (e.g., medical diagnosis) and finance (loan approval) were employed to evaluate fairness improvements. Metrics included causal path identification accuracy, fairness indices (path-specific counterfactual fairness), and decision accuracy. Ablation studies examined the impact of different explanation layers. Hyperparameters were tuned for robustness, demonstrating the model’s capacity to handle complex causal pathways and improve interpretability and fairness across scenarios.

The proposed causal explanation framework achieved approximately 15% higher accuracy in action-guiding tasks compared to feature attribution methods. It significantly reduced unfair influence by over 20% in path-specific fairness evaluations. The formal definitions of actual causation successfully isolated true causes in complex causal graphs, outperforming correlation-based approaches. Ablation results confirmed that layered explanations provide nuanced insights, with strong sufficiency explanations offering the best balance between interpretability and actionability.

Applicable in healthcare for causal diagnosis, in finance for fair lending decisions, and in policy-making for transparent interventions. The framework requires causal structure knowledge or reliable causal discovery methods. Its ability to identify true causes and support path-specific fairness makes it valuable for deploying trustworthy AI systems that need to explain not just what happened, but why and how to intervene effectively.

Dependence on complete causal models limits practical deployment where causal structures are uncertain or incomplete. Extending the framework to probabilistic models and dynamic systems remains future work. Computational complexity grows with system size, necessitating scalable algorithms. Handling unobserved confounders and causal discovery from data are additional challenges to address.