Fairness Measures for Regression Models via Gaussian Mixture Density Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for determining bias in regression machine learning models are inefficient in conserving computing resources and fail to accurately calculate group fairness measures, particularly for continuously distributed target and prediction variables, leading to erroneous results and resource wastage.
Innovation Solution
A bias detection system that estimates conditional densities using Gaussian mixtures to calculate independence, separation, and sufficiency measures, conserving resources by accurately determining fairness metrics through mutual information and entropy calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current systems are used to determine bias in regression machine learning models, then computing resources are consumed, but accuracy of fairness measures is poor and resources are wasted
Solution Approach 1:
The patent transforms the continuous target and prediction variables into discrete binned variables, changing the parameter representation from continuous to discrete. This enables the application of mutual information and entropy calculations that are computationally efficient while maintaining measurement accuracy for fairness assessment.
Solution Approach 2:
The patent replaces complex computational mechanisms with information-theoretic measures (mutual information and entropy). Instead of using resource-intensive bias detection algorithms, the system uses mathematically elegant information measures that provide accurate fairness assessment with reduced computational overhead.
2Reliability
If current systems calculate group fairness measures for continuously distributed variables, then computational complexity increases, but results are erroneous
Solution Approach 1:
The patent segments the continuous target and prediction variables into discrete bins, transforming them from continuous to categorical variables. This segmentation enables the use of mutual information and entropy calculations that are computationally simpler while producing reliable fairness measurement results.
Solution Approach 2:
The patent introduces mutual information and entropy as intermediary measures to assess fairness. These information-theoretic quantities serve as mediators that connect the binned variables to fairness assessment, providing a computationally efficient pathway from continuous inputs to reliable fairness outputs.
Data Source
AI summary
A device may receive sensitive attribute data, model prediction data, and true target data associated with a regression machine learning model, and may determine quantities of Gaussian components for Gaussian mixtures. The device may generate Gaussian mixtures of the quantities of Gaussian components based on the sensitive attribute data, the model prediction data, and the true target data, and may determine parameters of estimates of conditional densities by the Gaussian mixtures based on the sensitive attribute data, the model prediction data, and the true target data. The device may calculate an independence measure, a separation measure, and a sufficiency measure of the regression machine learning model based on the Gaussian mixtures and the parameters of the estimates of the conditional densities. The device may perform actions based on one or more of the independence measure, the separation measure, or the sufficiency measure.