Computing Actual Causes for Neural Network Predictions under Structured Causal Inputs
A research paper formalizes neural network explanations as Halpern-Pearl actual causes using Boolean Structural Causal Models. The method uses bound propagation and branch-and-bound to compute all minimal actual causes with completeness guarantees. It scales to search spaces up to 2.3×10^13 candidate pairs on SCMs with up to 28 nodes within 180 seconds per instance, outperforming brute-force, ILP, and heuristic search baselines. A case study shows that ignoring input dependencies inflates reported causes, with 14.9% being spurious under the SCM.
Researchers propose a method to explain neural network predictions by computing Halpern-Pearl actual causes under structured causal inputs modeled as Boolean Structural Causal Models. The approach uses bound propagation and branch-and-bound to efficiently find all minimal actual causes with formal completeness guarantees. Experiments demonstrate scalability to large search spaces and SCMs, outperforming existing baselines. A case study reveals that ignoring input dependencies leads to inflated and spurious explanations.
The method leverages bound propagation and branch-and-bound to prune the search space of candidate cause-contingency pairs, enabling exact computation of minimal HP causes in Boolean SCMs. This contrasts with prior work that either assumes feature independence or uses heuristic search without completeness guarantees. The approach scales to SCMs with up to 28 nodes and search spaces of 2.3×10^13 pairs, suggesting potential for real-world causal explanation tasks where input dependencies are known.
This research addresses a critical gap in trustworthy AI by providing causally grounded explanations for neural network decisions. As regulatory and enterprise demands for explainability grow, methods that account for input dependencies could become essential for high-stakes applications like healthcare, finance, and autonomous systems. The demonstrated scalability and formal guarantees may accelerate adoption in compliance-sensitive industries.
For enterprises deploying neural networks in regulated domains, this method offers a path to more accurate and trustworthy explanations, potentially reducing compliance risks and improving model debugging. The ability to identify spurious causes could lead to more robust models and better decision-making. However, practical adoption depends on the availability of causal models for input features, which may require domain expertise.
Next signals include extensions to non-Boolean SCMs, integration with popular XAI frameworks, and empirical validation on real-world datasets with known causal structures. The approach could also be combined with causal discovery methods to handle unknown input dependencies. Watch for follow-up work on efficient approximation algorithms for larger SCMs and applications in fairness and robustness auditing.