Explanatory Model Configuration for Consistent Black-Box AI Explanations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Explanatory models for black-box machine-learning models often produce inconsistent explanations when faced with slightly different inputs, leading to unintuitive and untrustworthy outputs.
Innovation Solution
A method involving one-hot encoding and synthetic data generation to configure an explanatory model, ensuring consistent explanations by tuning it with weighted linear regression based on outputs from a target model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If an explanatory model is used to explain black-box machine-learning models, then interpretability and understanding of decision-making processes are improved, but consistency and reliability of explanations across different inputs deteriorate
Solution Approach 1:
The explanation is segmented into local explanations for individual predictions and global explanations for overall model behavior. Local explanations focus on specific input instances while global explanations ensure consistency across the entire input space, resolving the contradiction between detailed interpretability and overall consistency.
Solution Approach 2:
The patent changes parameters of the explanatory model including sampling distributions, explanation methods (e.g., LIME, SHAP), and aggregation techniques to optimize both interpretability and consistency. By adjusting these parameters, the system achieves reliable consistent explanations while maintaining detailed interpretability where needed.
2Loss of information
If detailed explanations are provided for each input, then interpretability is improved, but computational complexity and processing time increase
Solution Approach 1:
The system performs preliminary actions by pre-computing global explanations and establishing consistent frameworks before processing individual inputs. This preliminary setup reduces the computational burden during actual explanation generation, maintaining detailed interpretability while lowering overall complexity.
Solution Approach 2:
The patent uses copying by generating explanations for representative input samples and reusing these explanation patterns across similar inputs. This approach maintains detailed interpretability for each input while significantly reducing computational complexity through pattern reuse rather than full recomputation.
3Reliability
If consistent explanations are enforced across all inputs, then reliability is improved, but adaptability to specific input nuances deteriorates
Solution Approach 1:
The patent applies local quality by allowing explanations to vary in detail and focus based on the specific input characteristics while maintaining overall consistency through a unified framework. Each input receives tailored explanations that adapt to its nuances while conforming to global consistency requirements, resolving the contradiction between reliability and adaptability.
Data Source
AI summary
In some examples, a system can receive input datasets that each include a respective set of variable values for the same set of variables, where the input datasets have different values for a particular variable in the set of variables. The system can generate encoded input datasets by encoding the different values of the particular variable. The system can also generate encoded background data. The system can then generate synthetic data based on the encoded input datasets and the encoded background data. The system can provide the synthetic data as input to a target model to receive outputs from the target model. The system can then configure an explanatory model based on the synthetic data and the outputs from the target model. The explanatory model can generate scores indicating how a selected set of values for the set of variables influences an output from the target model.