Disparity Mitigation in Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning algorithms in automated decisioning systems often introduce or perpetuate disparities in predictions between different classes of data, leading to unfair outcomes for minority or non-dominant classes.
Innovation Solution
A system and method that uses partial Jensen-Shannon divergence calculations to generate indiscernibility constraints, which are incorporated into a disparity-mitigating loss function to train machine learning models, ensuring fairness by aligning prediction distributions across classes during the training process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning algorithms are used in automated decisioning systems, then prediction accuracy and automation efficiency are improved, but disparity and unfairness between different data classes are introduced or perpetuated
Solution Approach 1:
The patent applies preliminary anti-action by introducing fairness constraints and disparity mitigation mechanisms during the model training phase. The system pre-establishes constraints that actively counteract potential disparity before it manifests in predictions, using techniques like adversarial training and fairness-aware loss functions to prevent unfair outcomes from occurring in the first place
Solution Approach 2:
The patent implements feedback mechanisms by continuously monitoring prediction distributions across different data classes and using this information to adjust training parameters. The system calculates disparity metrics from model outputs and feeds this information back into the training process to iteratively reduce fairness violations while maintaining accuracy
2Measurement precision
If machine learning models are trained to optimize accuracy, then prediction performance is improved, but fairness and indiscernibility between distinct classes deteriorate
Solution Approach 1:
The patent applies parameter changes by modifying the loss function parameters to include fairness constraints alongside accuracy objectives. The system adjusts training parameters such as fairness weighting factors and constraint thresholds to balance accuracy and fairness, transforming the optimization problem to simultaneously consider both metrics rather than prioritizing accuracy alone
Solution Approach 2:
The patent implements multi-functionality by designing a unified training framework that simultaneously optimizes for multiple objectives: accuracy, fairness, and indiscernibility between classes. The loss function and training process serve multiple purposes at once, making the model adaptable to diverse fairness requirements across different application contexts
3Adaptability or versatility
If disparity mitigation constraints are added to the training process, then fairness between classes is improved, but training complexity and computational requirements increase
Solution Approach 1:
The patent applies partial action by implementing fairness constraints selectively on specific model layers or specific data classes rather than uniformly across the entire model. The system applies disparity mitigation to the extent necessary to achieve fairness goals without over-constraining the model, reducing unnecessary computational overhead while maintaining fairness improvements
Data Source
AI summary
A system and method includes generating approximate distributions for distinct classes of data samples; computing a first partial Jensen-Shannon (JS) divergence and a second partial JS divergence based on the approximate distribution of the disparity affected class of data samples with reference to the approximate distribution of the control class of data samples; computing a disparity divergence based on the first partial JS divergence and the second partial JS divergence; generating a distribution-matching term based on the disparity divergence, wherein the distribution-matching term mitigates an inferential disparity between the control class of data samples and the disparity affected class of data samples during a training of an unconstrained artificial neural network; constructing a disparity-constrained loss function based on augmenting a target loss function with the distribution-matching term; and transforming the unconstrained ANN to a disparity-constrained ANN based on a training of the unconstrained ANN using the disparity-constrained loss function.