ML Model Parity Metric Optimization for Bias Mitigation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models trained on biased datasets produce biased predictions, making it difficult to identify and optimize for biases that affect overall performance, particularly in areas like credit and loan applications where features like race or gender are not explicitly included but influence outcomes.
Innovation Solution
A system and method that optimize machine learning models by creating slices of predictions based on input feature vectors, determining sensitive and base metrics, and calculating parity metrics to address biases, using metrics such as Recall parity, False Positive Rate parity, and Disparate Impact, allowing for the optimization of the model based on these parity metrics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If overall aggregate performance metrics are used to optimize the machine learning model, then the model's overall performance is improved, but the ability to identify and address biases is lost
Solution Approach 1:
The patent segments the predictions into multiple slices based on different feature values (e.g., demographic groups, geographic regions, time periods). This segmentation allows the system to calculate parity metrics for each slice separately, enabling identification of biases that would be masked in aggregate performance metrics. The segmentation transforms a single overall performance view into multiple granular views that reveal hidden disparities.
2Object-affected harmful factors
If the model is optimized based on parity metrics for sensitive groups, then bias mitigation is improved, but overall model performance may deteriorate
Solution Approach 1:
The patent applies partial optimization by focusing on specific slices that exhibit bias rather than uniformly optimizing all predictions. The system calculates parity metrics for each slice and selectively optimizes only those slices where bias is detected, leaving well-performing slices unchanged. This partial action approach mitigates bias while preserving overall model performance by avoiding unnecessary adjustments to already fair predictions.
Solution Approach 2:
The patent changes the optimization parameters from aggregate performance metrics to slice-specific parity metrics. By adjusting the evaluation criteria to consider demographic parity, equal opportunity, or other fairness metrics for different slices, the system reoptimizes the model to balance fairness and accuracy. This parameter change enables the model to meet both bias mitigation and performance requirements through multi-objective optimization.
3Measurement precision
If multiple slice-based parity metrics are calculated and optimized, then bias identification and mitigation capability is improved, but system complexity increases
Solution Approach 1:
The patent segments the optimization process into modular components: slice creation, metric calculation, bias detection, and selective optimization. Each component handles a specific aspect of the fairness assessment, making the overall complex system manageable through clear separation of concerns. This modular segmentation allows the system to handle multiple parity metrics without becoming unmanageably complex.
Solution Approach 2:
The patent implements a universal optimization framework that can handle multiple types of parity metrics (demographic parity, equal opportunity, predictive parity) and multiple slice dimensions (demographics, geography, time) through a single unified system. This multi-functional approach avoids the need for separate optimization systems for each fairness criterion, reducing overall system complexity while maintaining comprehensive bias detection and mitigation capabilities.
Data Source
AI summary
Techniques for optimizing a machine learning model. The techniques may include obtaining multiple predictions from a machine learning model, the predictions being based on at least one input feature vector, each input feature vector having one or more vector values; creating at least one slice of the predictions based on at least one vector value; determining a sensitive bias metric for the slice based on a sensitive group; determining a base metric for the slice based on a base group; determining a parity metric for the slice based on a ratio of the sensitive bias metric and the base metric; and optimizing the machine learning model based on the parity metric.