Classifier Prediction Adjustment for Fairness and Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning classifier models face challenges in balancing classification accuracy with fairness, particularly in ensuring unbiased predictions across different population groups, and current techniques often require access to sensitive attributes during inference, which is impractical in real-world applications due to privacy concerns.
Innovation Solution
A post-processing method using bias scores to modify classifier predictions based on instance-level fairness criteria, such as Demographic Parity (DP), Equalized Opportunity (EOp), and Equalized Odds (EO), without requiring access to sensitive attributes during inference, by deriving a Bayes-optimal classifier and applying a modification rule to adjust predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If fairness constraints are applied during classifier training, then prediction fairness across population groups is improved, but classification accuracy deteriorates
Solution Approach 1:
The patent pre-computes bias scores for different population groups during the training phase, storing these scores in a lookup table. During inference, the system retrieves pre-computed bias scores rather than recalculating them, enabling rapid adjustment of predictions to satisfy fairness constraints without compromising classification accuracy. This preliminary preparation resolves the contradiction by decoupling the computationally intensive fairness analysis from the real-time prediction process.
Solution Approach 2:
The patent introduces bias scores as an intermediary metric that quantifies the disparity in prediction outcomes across different population groups. These bias scores serve as a mediator between the classifier's raw predictions and the final adjusted predictions, allowing the system to enforce fairness constraints through a controlled transformation process that preserves overall classification accuracy while correcting group-specific biases.
2Quantity of substance
If techniques requiring access to sensitive attributes during inference are used, then fairness constraints can be satisfied, but privacy protection and practical deployment are compromised
Solution Approach 1:
The patent pre-computes bias scores for different population groups during the training phase, storing these scores in a lookup table. During inference, the system retrieves pre-computed bias scores rather than recalculating them, enabling rapid adjustment of predictions to satisfy fairness constraints without compromising classification accuracy. This preliminary preparation resolves the contradiction by decoupling the computationally intensive fairness analysis from the real-time prediction process.
Solution Approach 2:
The system uses the classifier's own training data and learned representations to compute bias scores during training, making the fairness adjustment mechanism self-contained. The bias scores are derived from the classifier's internal decision boundaries and training distribution, eliminating the need for external sensitive attribute information during inference while still enabling effective fairness enforcement through the pre-computed scores.
3Device complexity
If the trade-off between accuracy and fairness is fixed during training, then model simplicity is maintained, but flexibility in adjusting fairness levels deteriorates
Solution Approach 1:
The patent implements a dynamic fairness adjustment mechanism where the bias correction amount can be controlled through a parameter alpha that ranges from 0 to 1. This allows the system to adaptively balance between accuracy and fairness based on different application requirements, transforming the static fairness-accuracy trade-off into a controllable spectrum of options without increasing fundamental model complexity.
Solution Approach 2:
The patent introduces a controllable parameter alpha that modulates the degree of bias correction applied to predictions. By varying this parameter, users can dynamically adjust the fairness level of the classifier output, enabling flexible adaptation to different deployment scenarios while maintaining a relatively simple model architecture. This parameter-based control resolves the contradiction by providing versatility without proportional increases in complexity.
Data Source
AI summary
The present disclosure describes techniques for balancing classification accuracy and fairness of a model trained to perform classification tasks. At least one bias score function corresponding to each sensitive attribute associated with instances classified by the model is configured. The at least one bias score function is configured to measure fairness on an instance level. At least one modification rule is generated based on the at least one bias score function and parameters. The at least one modification rule corresponds to at least one fairness criterion. The parameters are associated with a target level of the at least one fairness criterion. At least a subset of predictions are modified by applying the at least one modification rule to the predictions generated by the model. The modified predictions satisfy the target level of the at least one fairness criterion while maintaining the classification accuracy.