ML Bias Detection Using Proxy Protected Attributes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models often contain latent bias, which is difficult to detect, especially when protected attributes are not included as features or known in the training data, posing challenges in ensuring fairness and accuracy.
Innovation Solution
A method for automated detection of bias in machine learning models by predicting protected attributes based on identifying information such as names and geographic locations, and using statistical parity, equal opportunity, and predictive equality to analyze model outputs for different classes, generating a bias report to identify and address any detected bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If protected attributes are excluded from training data to prevent explicit discrimination, then fairness is improved, but latent bias detection becomes difficult
Solution Approach 1:
The patent introduces proxy features as intermediaries that correlate with protected attributes without being the protected attributes themselves. These proxy features serve as mediators that allow the system to detect latent bias indirectly - the bias detection module analyzes correlations between proxy features and model outputs to identify hidden discrimination patterns even when protected attributes are excluded from training
Solution Approach 2:
The patent implements a feedback mechanism where model outputs are analyzed against proxy feature distributions to detect bias patterns. The bias detection module provides feedback about detected latent bias to the training process, enabling iterative improvement of model fairness while maintaining the exclusion of protected attributes from training data
2Measurement precision
If test data set size is increased to produce valuable bias detection results, then measurement precision is improved, but data requirements and complexity increase
Solution Approach 1:
The patent applies partial action by using synthetic proxy feature data generated from statistical distributions rather than requiring complete protected attribute information for all test instances. This approach provides sufficient bias detection capability without the excessive data collection and processing complexity that would be required to obtain actual protected attributes for large-scale testing
Data Source
AI summary
Aspects of the present disclosure provide techniques for detecting latent bias in machine learning models. Embodiments include receiving a data set comprising features of a plurality of individuals. Embodiments include receiving identifying information for each individual of the plurality of individuals. Embodiments include predicting, for each respective individual of the plurality of individuals, a probability that the respective individual belongs to a given class based on the identifying information for the given individual. Embodiments include providing, as inputs to a machine learning model, the features of the plurality of individuals from the data set. Embodiments include receiving outputs from the machine learning model in response to the inputs. Embodiments include determining whether the machine learning model is biased against the given class based on the outputs and the probability that each respective individual of the plurality of individuals belongs to the given class.