Fairness-Enhanced Machine Learning Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine-learning models in the financial industry often lack fairness, resulting in disparate treatment or impact on legally protected groups, posing legal issues and performance challenges when attempting to address these concerns through manual data removal.
Innovation Solution
Incorporating fairness improvements into the model training process by using control parameters to minimize disparities, enabling self-adaptive control vectors that balance model performance and fairness, and automatically generating fair data labeling to develop models that satisfy specific fairness metrics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine-learning models are trained to maximize prediction power, then model accuracy is improved, but fairness deteriorates resulting in disparate treatment or impact on protected groups
Solution Approach 1:
The patent transforms the fairness improvement problem from a post-processing step to an integrated part of the model training process by changing the optimization parameters. It introduces control parameters (lambda vectors) that allow simultaneous optimization of both prediction power and fairness metrics during training, rather than treating them as conflicting objectives requiring separate handling.
Solution Approach 2:
The patent performs preliminary action by automatically generating fair data labeling and incorporating fairness constraints before the actual model training occurs. The system pre-processes the training data to ensure fairness requirements are embedded in the data itself, and pre-defines control parameters that will guide the optimization process to balance accuracy and fairness from the outset.
2Object-affected harmful factors
If manual data removal is used to address fairness concerns, then disparate impact is reduced, but model performance deteriorates
Solution Approach 1:
The patent implements feedback mechanisms where the system continuously monitors both fairness metrics and model performance during training. The control parameters are adjusted based on feedback from fairness evaluations, allowing the model to learn from discrepancies and improve both fairness and performance iteratively rather than through static data removal.
Solution Approach 2:
The system performs self-service by automatically generating fair data labeling and adjusting control parameters without requiring manual intervention. The machine-learning model itself is equipped with the capability to identify and correct its own fairness issues through the integrated optimization framework, eliminating the need for external manual data processing.
3Object-affected harmful factors
If fairness metrics are enforced during training, then fairness is improved, but training complexity increases
Solution Approach 1:
The patent merges the fairness optimization objective with the traditional prediction accuracy objective into a single unified loss function. By combining these objectives with appropriate weighting through control parameters, the system handles multiple goals simultaneously in one optimization process rather than requiring separate training runs or complex multi-objective optimization frameworks.
Data Source
AI summary
A method including training a machine-learning model, based on historical data, with a maximization problem and one or more minimization problems to improve one or more fairness metrics. The method also can include receiving real-time data. The method additionally can include generating a risk score based on the machine-learning model, as trained, and the real-time data. Other embodiments are described.