The paper proposes BayesSubX, a framework that combines Bayesian optimization with submodular function maximization to achieve robust explainability in AI. It treats feature selection for explanations as a sequential decision problem under uncertainty, using Gaussian Process surrogates to model feature importance and submodular constraints for diversity. The framework introduces probabilistic acquisition functions for balancing exploration and exploitation, providing inherent robustness through uncertainty quantification.
Key findings
Proposes BayesSubX, a novel framework for robust explainability in AI.
Combines Bayesian optimization with submodular function maximization.
Uses Gaussian Process surrogates to model feature importance landscape.
Introduces probabilistic acquisition functions for balancing exploration and exploitation.
Provides theoretical guarantees and comprehensive experimental validation.
Limitations & open questions
The framework may have limitations in handling very high-dimensional feature spaces.
The computational cost of Gaussian Process models may be prohibitive for large-scale problems.
Further research is needed to extend the approach to other types of explainable AI models.