Optimal Transport Data Preprocessing for Algorithmic Fairness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for achieving algorithmic fairness in machine learning models often require modifying the original data, which can be unethical or costly, especially in high-stakes settings like finance and healthcare.
Innovation Solution
The method involves using optimal transport to preprocess data by computing sample-level weights that minimize the Wasserstein distance between the original dataset and a weighted version, while satisfying demographic parity constraints, thus reducing disparities in classification datasets without modifying the original data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If pre-processing methods change feature values or labels of training data, then fairness disparities are reduced, but data integrity and ethical concerns are compromised
Solution Approach 1:
The patent introduces sample weights as an intermediary mechanism that adjusts model training without modifying the original data. Instead of directly changing feature values or labels, the method uses weight parameters that mediate between the original data and the learning algorithm, achieving fairness goals while preserving data integrity.
Solution Approach 2:
The patent changes the parameter space from modifying data values to adjusting sample weights. By transforming the problem from data manipulation to parameter optimization, the method achieves fairness through weight adjustments rather than unethical data modification.
2Reliability
If synthetic or modified versions of datasets are created, then fairness is improved, but storage costs and complexity increase
Solution Approach 1:
Instead of creating full synthetic copies of datasets, the patent uses weight parameters that virtually replicate the effect of modified data. The weights act as a compact representation that encodes fairness adjustments without requiring physical duplication of data.
Solution Approach 2:
The patent transforms the problem from creating additional data copies to optimizing parameter weights. This parameter-based approach achieves the same fairness effect with minimal storage overhead, avoiding the need to maintain separate synthetic datasets.
3Reliability
If sample weights are learned to satisfy fairness constraints, then demographic parity is improved, but guarantees about data distribution alteration are lost
Solution Approach 1:
The patent implements a feedback mechanism where sample weights are iteratively adjusted based on measured demographic parity violations. The system continuously monitors fairness metrics and refines weights to satisfy constraints, providing both fairness guarantees and controlled distribution alteration.
Solution Approach 2:
The patent performs preliminary analysis of data distribution and fairness constraints before learning weights. By pre-characterizing the data and constraints, the method establishes bounds and guarantees about distribution alteration before the actual weight learning process begins.
Data Source
AI summary
A method and a system for pre-processing data for algorithmic fairness via optimal transport in order to reduce disparities in classification datasets without modifying the original data are provided. The method includes: receiving a first dataset that includes a set of samples, each respective sample including a first coordinate that relates to sensitive demographic features, a second coordinate that relates to decision-making features, and a third coordinate that relates to a decision outcome that is generated by a machine learning model; determining a demographic parity constraint to be applied to the first dataset; computing a set of respective sample-level weights that correspond to each sample; and generating a second dataset by applying the set of respective sample-level weights to the first dataset The computation of the sample-level weights includes minimizing a Wasserstein distance between the first dataset and a weighted version thereof while satisfying the demographic parity constraint.