Global Model Explainability for Accuracy-Preserving Data Minimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Determining what data is strictly necessary for an ML model to maintain a threshold level of accuracy is challenging, particularly in compliance with regulations like GDPR, which requires data minimization.
Innovation Solution
Utilizing global explainability analysis techniques like SHAP and LIME to identify features and values that have minimal influence on model accuracy, allowing for data generalization by replacing precise values with generalized ranges or groups, while maintaining or slightly reducing the model's performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If precise feature values are collected and retained, then model accuracy is maintained, but data volume increases and compliance with data minimization regulations becomes difficult
Solution Approach 1:
The patent segments the feature space by identifying distinct regions or clusters of feature values that lead to similar model predictions. By grouping individual feature values into broader categories or ranges, the system reduces the total number of distinct values stored while preserving predictive accuracy. This segmentation allows data minimization by replacing granular feature values with coarser representations.
Solution Approach 2:
The patent transforms feature values by applying parameter changes such as discretization, binning, or mapping continuous values to categorical representations. This transformation reduces the precision of individual feature values while maintaining the overall information needed for accurate predictions, thereby reducing data volume without significantly compromising model reliability.
2Quantity of substance
If data is generalized to reduce volume, then data minimization and compliance improve, but model accuracy may deteriorate
Solution Approach 1:
The patent employs feedback mechanisms where the model's performance is continuously monitored after generalization is applied. If accuracy degradation is detected, the system adjusts the generalization parameters or selects different feature values to generalize, ensuring that model performance remains within acceptable thresholds. This feedback loop enables iterative optimization of the generalization process.
Solution Approach 2:
The patent applies partial generalization selectively to only those feature values that have minimal impact on model accuracy. By identifying and generalizing only the least critical features or value ranges, the system achieves data minimization while preserving the accuracy-critical portions of the data in their original form.
3Loss of information
If all data is collected indefinitely, then complete information is available for analysis, but compliance with retention limits and privacy regulations is violated
Solution Approach 1:
The patent creates simplified copies or representations of the original data through generalization. Instead of storing and processing all raw feature values, the system stores generalized versions that capture the essential patterns needed for predictions. This copying approach reduces data retention requirements while maintaining analytical utility.
Solution Approach 2:
The patent extracts only the essential information needed for model predictions by identifying and retaining only the most influential feature values or ranges. Less critical data is either generalized or discarded, extracting the core predictive signal while removing excess information that contributes to data volume but not to predictive accuracy.
Data Source
AI summary
An embodiment analyzes a predictive model and its input data for the predictive model using an explainability algorithm resulting in a feature importance value of a feature. The embodiment analyzes feature values of the feature using a generalization function resulting in a set of candidate feature values. The embodiment determines an alternative feature based on the set of candidate feature values, wherein the alternative feature is a generalization of the feature. The embodiment compares an accuracy of the predictive model to a threshold performance value and, responsive to the accuracy being above the threshold performance value, maps feature values in the input data that are in the set of candidate feature values to a generalized representative value in the generalized domain.