Certifying Black Box ML Input Subspaces for Stable Explanations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for explaining black box machine learning models lack guarantees of stability and wide applicability, as they typically require individual explanations for each input example, without providing certification of the validity of these explanations across a region.
Innovation Solution
The method involves certifying a d-dimensional input space for a black box machine learning model by identifying the largest subspace where the explanation remains valid, using a quality metric and a specified quality threshold, and executing the model within this certified subspace to improve its performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If individual explanations are provided for each input example, then explanation accuracy is improved, but explanation stability and wide applicability deteriorate
Solution Approach 1:
The patent segments the input space into a certified region (where explanations are valid) and uncertified regions. By identifying and isolating the certified region through subspace certification, the method ensures that explanations remain stable and applicable across the entire certified region rather than being valid only for individual points.
Solution Approach 2:
The patent changes the parameter from point-specific explanations to region-wide explanations by introducing a quality threshold θ and certifying a maximum subspace. This transforms explanations from being accurate at individual points to being stable across a certified region, resolving the contradiction between precision and stability.
2Adaptability or versatility
If explanations are certified across a larger region, then explanation applicability is improved, but computational cost increases
Solution Approach 1:
The patent applies partial action by certifying only the maximum subspace necessary for explanation validity rather than the entire input space. The Ecertify process stops when the quality threshold θ is no longer satisfied, avoiding unnecessary computational effort on regions where explanations would not be valid.
Solution Approach 2:
The patent replaces exhaustive computational search with a more efficient certification approach using random sampling and quality metric evaluation. Instead of systematically checking all possible input points, the method uses stochastic sampling to identify the certified region boundaries, reducing computational cost while maintaining applicability.
3Productivity
If random sampling is used to certify the maximum subspace, then computational efficiency is improved, but measurement precision of the certified region boundaries deteriorates
Solution Approach 1:
The patent implements feedback through iterative refinement of the certified region. The Ecertify process repeatedly samples points, evaluates quality metrics, and adjusts the certified subspace boundaries based on whether points satisfy the quality threshold. This feedback loop continues until convergence, ensuring precise boundaries are identified through stochastic processes.
Solution Approach 2:
The patent applies preliminary action by establishing a quality threshold θ and sampling framework before final region boundary determination. This preliminary setup enables efficient random sampling to converge on precise boundaries, as the sampling process is guided by the pre-defined quality criteria rather than requiring exhaustive search.
Data Source
AI summary
A method, computer program product, and computer system for certifying a d-dimensional input space x for a black box machine learning model. Triggered is execution of a first process that certifies, with respect to the model, a maximum subspace of x that is characterized by a largest half-width or radius (w) centered at x=x0. Received from of the first process are: w and both (i) a point re selected from multiple points r randomly sampled in the maximum subspace, and (ii) a quality metric f(re), where re and f(re) were previously determined from the model having been queried for each point r randomly sampled in the maximum subspace, where re is selected on a basis of f(re) satisfying f(re)≥θ for a specified quality threshold θ. The model is executed for input confined to the maximum subspace, which performs a practical application procedure that improves performance of the model.