A Computationally Feasible Framework for Causal Probabilistic Explanation
The paper introduces Probabilistic Causal Impact (PCI), a framework that builds on actual causality and Pearl's notions of probability of necessity and sufficiency. PCI recasts explainability as an estimation problem on a probabilistic causal model, approximated via Monte Carlo, providing tractable, causally grounded, graded explanations. It generalizes actual causality and Pearl's probability of causation as degenerate cases. The framework is evaluated on synthetic and real-world examples.
Researchers propose Probabilistic Causal Impact (PCI), a new framework for causal explanation that bridges the gap between principled but computationally intractable actual causality methods and scalable but causally naive attribution methods like SHAP. PCI formulates explanation as an estimation problem over a probabilistic causal model, using Monte Carlo approximation to deliver graded, causally grounded explanations. It generalizes both actual causality and Pearl's probability of causation, and is demonstrated on synthetic and real-world examples.
PCI shifts from exact counterfactual enumeration to Monte Carlo estimation over distributions of candidate explanations and counterfactual values, making causal explanation computationally feasible for larger models. This approach preserves causal grounding while allowing graded, probabilistic attributions, potentially enabling integration with modern probabilistic programming and simulation-based inference.
The framework could improve explainability in high-stakes AI applications where causal structure matters, such as healthcare, finance, and policy analysis. By offering tractable causal explanations, PCI may reduce reliance on SHAP-like methods that can produce causally inconsistent attributions, addressing a key limitation in current model interpretability tooling.
PCI offers a path to more trustworthy and causally valid model explanations, which could differentiate AI products in regulated industries and improve debugging of causal models. It may lower the cost of obtaining principled explanations, enabling wider deployment of causal inference in enterprise analytics.
Next observable signals include open-source implementations of PCI, benchmarks comparing PCI against SHAP and causal SHAP on standard datasets, and adoption in probabilistic programming libraries. Further research may extend PCI to handle latent confounders, high-dimensional data, and integration with deep generative models.