Fair Machine Learning Training via Iterative Bound Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models often exhibit bias, particularly in critical applications, due to historical disparities and biased training data, leading to unfair treatment of certain groups based on attributes like ethnicity, age, or gender, necessitating the development of fair prediction models that calibrate predictions across different demographics.
Innovation Solution
A method involving a non-transitory computer-readable medium with instructions to train a fair machine learning model by initializing a bound value, weighting observation vectors based on fairness constraints, and iteratively updating the bound value to achieve equalized odds fairness, ensuring the model predicts target variables without bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning models are trained using historical data, then productivity and cost savings are improved, but bias and unfair treatment of certain groups are introduced
Solution Approach 1:
The patent applies preliminary action by performing fairness-aware reweighting of training observations before model training. The system calculates importance weights for each training example based on group membership and prediction accuracy, then uses these weights to retrain the model. This preliminary preprocessing step ensures that the model learns from a balanced representation of different groups, preventing bias from being encoded into the decision-making process while still utilizing historical data for productivity gains.
2Measurement precision
If machine learning models are trained to improve accuracy, then prediction quality is improved, but bias against certain demographics is amplified
Solution Approach 1:
The patent applies parameter changes by modifying the training objective through importance weights. Instead of treating all training examples equally, the system changes the effective parameter of each example by assigning different weights based on group representation and prediction performance. This parameter transformation allows the model to maintain high overall accuracy while specifically improving accuracy for underrepresented groups, thereby reducing demographic bias without sacrificing prediction quality.
3Object-affected harmful factors
If fairness constraints are applied during model training, then bias is reduced, but training complexity and computational resources increase
Solution Approach 1:
The patent applies segmentation by dividing the fairness constraint problem into distinct computational stages. First, the system segments the training data into groups based on sensitive attributes. Second, it segments the computation by calculating importance weights in a separate preprocessing step before model training. This segmentation transforms a complex constrained optimization problem into a simpler two-stage process: weight calculation followed by standard weighted training, significantly reducing training complexity while maintaining bias reduction effectiveness.
Data Source
AI summary
A computing device trains a fair machine learning model. A predicted target variable is defined using a trained prediction model. The prediction model is trained with weighted observation vectors. The predicted target variable is updated using the prediction model trained with weighted observation vectors. A true conditional moments matrix and a false conditional moments matrix are computed. The training and updating with weighted observation vectors are repeated until a number of iterations is performed. When a computed conditional moments matrix indicates to adjust a bound value, the bound value is updated based on an upper bound value or a lower bound value, and the repeated training and updating with weighted observation vectors is repeated with the bound value replaced with the updated bound value until the conditional moments matrix indicates no further adjustment of the bound value is needed. A fair prediction model is trained with the updated bound value.