Verified SHAP: Provable Bounds for Exact Shapley Values of Neural Networks
An innovative algorithm utilizes neural network verification to determine precise lower and upper bounds on SHAP values, effectively retrieving exact SHAP values. This method is capable of scaling to significantly larger search spaces compared to leading exact techniques, marking an initial advancement towards precise SHAP calculations. Additionally, it serves as a foundational element for assessing statistical approximation methods within broader search environments.
Key facts
- Shapley additive explanations (SHAP) are widely recognised as computationally intractable for neural networks.
- SHAP induces an exponential search space over input features.
- The algorithm leverages recent advances in neural network verification.
- It computes arbitrarily tight exact lower and upper bounds on SHAP values.
- It ultimately recovers the exact SHAP values.
- The approach scales to orders of magnitude larger search spaces than state-of-the-art exact methods.
- This provides an important first step towards exact SHAP computation.
- It establishes a principled cornerstone for evaluating statistical approximation methods on larger search spaces.
Entities
Institutions
- arXiv