Machine Learning Model Explanation Porting via Mediator
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning model repair methods focus on predictive metrics, data augmentation, and fairness, but lack effective techniques for importing desirable explanations from one model to another, especially for improving explainability and user trust.
Innovation Solution
A method that identifies a sample data point with differing explanations from two models, generates candidate samples, selects a subset based on explanation differences, and re-trains the first model using these samples to align its explanations with the target explanation from the second model, thereby enhancing explainability and user trust.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional model repair methods focus on predictive metrics and data augmentation, then model accuracy is improved, but model explainability remains insufficient
Solution Approach 1:
The patent uses a third model (model 304) as an intermediary to generate target explanations that are then imported into the first model. This mediator model serves as a bridge between the black-box model and the desired explanation format, allowing the transfer of explainability without compromising the original model's predictive accuracy.
Solution Approach 2:
The patent segments the model repair process into distinct components: generating candidate explanations, selecting target explanations, generating candidate samples, and retraining the model. This segmentation allows each component to be optimized independently, with the explanation import mechanism working separately from the accuracy optimization processes.
2Ease of operation
If model explanations are altered to improve understandability, then user trust is enhanced, but model prediction accuracy may deteriorate
Solution Approach 1:
The patent applies local quality by generating candidate samples within a specified neighborhood of the source sample rather than modifying the entire training dataset. This localized approach ensures that explanation improvements are concentrated in the relevant decision region while preserving the global prediction accuracy of the model.
Solution Approach 2:
The patent uses partial action by selecting only a subset of candidate samples for retraining rather than using all generated samples. This selective approach applies just enough modification to improve explanations while avoiding excessive changes that would degrade prediction accuracy. The subset is chosen based on criteria that balance explanation quality and accuracy preservation.
Data Source
AI summary
Retraining a model to present a target explanation with a prediction responsive to a source sample. The target explanation being selected from explanations provided by at least two machine learning models. A set of candidate samples is selected from samples generated from a relationship to the source sample. The retaining being performed with the set of candidate samples in a revised training dataset and causing a model presenting another explanation to present the target explanation with the prediction responsive to the source sample.