Fair ML Training Using Sensitive Subspaces and Robust Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models face challenges in ensuring individual fairness, as group fairness definitions are mutually incompatible and lack consensus on similarity among users, leading to unfairness from individual perspectives.
Innovation Solution
The method involves constructing a sensitive subspace using fair metrics and employing distributionally robust optimization to train machine learning models, ensuring individual fairness by minimizing perturbations along sensitive dimensions while maintaining accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If group fairness definitions are used for machine learning models, then statistical analysis is amenable, but definitions are mutually incompatible and algorithms may be unfair from individual users' perspectives
Solution Approach 1:
The patent segments the feature space by constructing a sensitive subspace that separates protected attributes from other features. This segmentation allows the model to apply different fairness constraints to different subspaces, enabling both group fairness (through subspace constraints) and individual fairness (through metric learning on the remaining features) to be satisfied simultaneously.
Solution Approach 2:
The patent introduces a new dimension by learning a custom distance metric in the non-sensitive subspace. This metric learning approach adds a dimensional transformation that enables individual fairness by ensuring similar treatment for similar individuals, while the orthogonal sensitive subspace maintains group fairness through independent constraint application.
2Reliability
If individual fairness is enforced in machine learning models, then fairness for similar users is improved, but there is a lack of consensus regarding which users are similar for ML tasks
Solution Approach 1:
The patent introduces a sensitive subspace as an intermediary that mediates between protected attributes and the main feature space. By projecting features onto this subspace and learning a distance metric on the orthogonal complement, the approach provides a systematic way to define similarity that respects individual fairness without requiring ad-hoc definitions for each ML task.
Solution Approach 2:
The patent changes the parameter space by learning a custom distance metric rather than using standard Euclidean distance. This parameter transformation allows the model to adapt the notion of similarity to the specific ML task while enforcing individual fairness, resolving the consensus problem by making similarity task-specific rather than universal.
3Reliability
If distributionally robust optimization is used to train machine learning models, then individual fairness is achieved, but training complexity and computational requirements increase
Solution Approach 1:
The patent segments the optimization problem into two parts: (1) constructing the sensitive subspace through principal component analysis or similar dimensionality reduction, and (2) applying distributionally robust optimization only on the reduced non-sensitive subspace. This segmentation reduces computational complexity while maintaining individual fairness guarantees.
Solution Approach 2:
The patent applies distributionally robust optimization partially, only to the non-sensitive subspace rather than the entire feature space. This partial application reduces computational burden while still achieving individual fairness for the relevant features, avoiding the excessive complexity of full-space DRO.
Data Source
AI summary
Obtain a first data set, a second data set, and a machine learning model. Construct a sensitive subspace of the first data set that defines a fair metric for distance among elements of the first data set. Fairly train the machine learning model on the first data set using a distributionally robust optimization approach based on the fair metric. Produce an individually fair set of labels by applying the fairly trained machine learning model to the second data set. Allocate a resource according to the individually fair set of labels.