Machine Learning Bias Estimation Without Losing Predictive Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Supervised learning models trained on historical data with discriminatory features can replicate biases present in the training data, leading to unfair outcomes, and removing these features may reduce predictive accuracy and fail to address information leakage.
Innovation Solution
A method to identify discriminatory identifier features through secondary models, analyze their influence on the main model, and estimate a discrimination likelihood to provide insights for model evaluation and improve predictive accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If discriminatory features are removed from training data records, then bias reduction is improved, but predictive accuracy deteriorates
Solution Approach 1:
The patent introduces an intermediary evaluation framework that assesses the impact of removing discriminatory features through controlled experiments and statistical analysis. This intermediary layer allows researchers to measure both bias reduction and predictive accuracy without directly modifying the training data, thus resolving the contradiction between these two objectives.
Solution Approach 2:
The patent changes the parameter of feature selection by systematically removing discriminatory features while compensating through adjusted weighting schemes and retraining procedures. This parameter change approach allows the model to maintain predictive accuracy while reducing bias, as the removal is compensated by optimized feature weighting and multiple training iterations.
2Object-affected harmful factors
If discriminatory features are removed from training data records, then bias reduction is improved, but information leakage persists
Solution Approach 1:
The patent implements a feedback mechanism that continuously monitors for information leakage by evaluating model predictions against known discriminatory patterns. This feedback loop allows the system to detect and adjust for lingering biases and information leakage, ensuring that removing discriminatory features does not create new problems.
Solution Approach 2:
The patent replaces direct mechanical removal of discriminatory features with a statistical and algorithmic approach that substitutes feature removal with feature transformation and weighting adjustments. This substitution maintains information integrity while eliminating bias, as the system replaces simple deletion with intelligent feature manipulation that preserves necessary information without leakage.
3Ease of manufacture
If feature removal is applied to address bias, then implementation simplicity is improved, but effectiveness deteriorates
Solution Approach 1:
The patent segments the bias reduction process into distinct, manageable steps: identifying discriminatory features, removing them from training data, evaluating impact through controlled experiments, and adjusting model parameters. This segmentation makes the complex task of bias reduction more implementable while maintaining effectiveness, as each step can be independently optimized and monitored.
Data Source
AI summary
Systems and methods include reception of a plurality of records, each of the plurality of records associating each of a plurality of features with a respective value, a second feature with a value, and a target feature with a value, a first machine learning model trained based on the plurality of records to output a value of the target feature based on values of each of the plurality of features, a second machine learning model trained based on the plurality of records to output a value of the second feature based on the values of each of the plurality of features, determination, based on the trained second machine learning model, of a first one or more of the plurality of features which are correlated to the second feature, determination of an influence of each of the first one or more features on the trained first machine learning model, and determination of a first value associated with the second feature based on the determined influences and on the trained second machine learning model.