Fairness Feature Importance Mitigating Unjustifiable Bias in ML Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current bias mitigation algorithms in machine learning models cannot distinguish between justifiable and unfair bias, leading to unnecessary decreases in model accuracy when attempting to improve fairness metrics.
Innovation Solution
An interactive feature selection method that measures fairness impact and accuracy impact separately, allowing users to identify and retain justifiable features, thereby increasing model accuracy while maintaining fairness by using fairness impact measurement for sensitive features that can be permuted.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If bias mitigation algorithms are applied to improve fairness metrics, then fairness is improved, but model accuracy decreases unnecessarily
Solution Approach 1:
The patent segments bias into two distinct categories: justifiable bias and unjustifiable bias. Justifiable bias refers to bias that arises from using features that are logically sound predictors of the outcome, while unjustifiable bias refers to bias that arises from using features that are correlated with protected attributes but do not provide legitimate predictive value. By segmenting bias in this way, the patent enables selective mitigation of only the unjustifiable portion, preserving the accuracy-contributing justifiable bias.
Solution Approach 2:
The patent introduces an intermediary mechanism - a analytical framework that measures and evaluates bias to distinguish between justifiable and unjustifiable types. This intermediary assessment layer sits between the raw model predictions and the final bias mitigation action, allowing for nuanced differentiation and selective intervention. The framework acts as a mediator that determines which bias should be addressed and which should be preserved.
2Reliability
If features correlated with protected attributes are removed to improve fairness, then fairness is improved, but model accuracy decreases
Solution Approach 1:
The patent applies local quality by treating different features differently based on their specific characteristics and relationship to the outcome. Rather than applying a blanket removal of all features correlated with protected attributes, the patent evaluates each feature individually to determine whether the bias it introduces is justifiable or unjustifiable. This localized assessment ensures that only features with unjustifiable bias are removed, while features with justifiable bias are retained.
3Reliability
If bias mitigation is applied to achieve fairness, then fairness metrics improve, but the distinction between justifiable and unjustifiable bias is lost
Solution Approach 1:
The patent implements feedback by creating a measurement and evaluation framework that provides information about the nature and source of bias in the model. This feedback mechanism allows users to understand whether bias is justifiable or unjustifiable, enabling informed decisions about which bias to mitigate. The feedback loop preserves information about bias justification while still allowing fairness improvement through selective mitigation.
Data Source
AI summary
In an embodiment, a computer generates a respective original inference from each of many records. Permuted values are selected for a feature from original values of the feature. Based on the permuted values for the feature, a permuted inference is generated from each record. Fairness and accuracy of the original and permuted inferences are measured. For each of many features, the computer measures a respective impact on fairness of a machine learning model, and a respective impact on accuracy of the machine learning model. A global explanation of the machine learning model is generated and presented based on, for multiple features, the impacts on fairness and accuracy. Based on the global explanation, an interactive indication to exclude or include a particular feature is received. The machine learning model is (re-)trained based on the interactive indication to exclude or include the particular feature, which may increase the fairness of the model.