Local Fairness Degree Determination for ML Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Machine Learning (ML) models lack the ability to assess local fairness, which is crucial for ensuring individual predictions are unbiased, as existing methods focus on global fairness and require computationally intensive approaches or domain expert input for synthetic data generation.
Innovation Solution
A method and system using Gaussian Mixture Model (GMM) clustering and constrained perturbation techniques to determine local fairness by generating perturbed datasets and calculating similarity scores, allowing for the evaluation of ML model decisions on individual instances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If localization-based test generation is used to generate perturbed data for individual fairness testing, then individual fairness insights are obtained, but computational complexity and time consumption increase significantly
Solution Approach 1:
The patent pre-computes and stores the covariance matrix and other statistical properties of the training data during an offline preparation phase. This preliminary action allows the online individual fairness testing to directly use these pre-computed values without recalculating them for each test instance, significantly reducing online computational time while maintaining measurement precision
Solution Approach 2:
The patent segments the fairness testing process into offline data preparation phase and online testing phase. The computationally intensive operations (covariance matrix computation, data generation) are performed offline and stored, while the online phase only requires simple similarity calculations and threshold comparisons, thus reducing real-time computational burden
2Measurement precision
If domain expert input is required to define thresholds for each dataset, then fairness verification accuracy improves, but system complexity and ease of operation deteriorate
Solution Approach 1:
The system automatically computes the similarity threshold based on the distribution of similarity scores between counterfactual instances and original instances in the training data. This self-service approach eliminates the need for domain experts to manually define thresholds, making the system easier to operate while maintaining statistically sound fairness verification
Solution Approach 2:
The patent transforms the threshold from a domain-expert-defined parameter into a data-driven parameter that is automatically computed from the training data distribution. The threshold is derived as a function of the similarity score distribution, adapting automatically to different datasets without requiring expert input for each new dataset
3Quantity of substance
If global fairness analysis is performed, then overall model behavior is understood, but individual prediction fairness insights are lost
Solution Approach 1:
The patent implements local fairness testing by generating counterfactual instances specific to each individual prediction and computing similarity scores locally. This allows the system to assess fairness for each individual decision while maintaining the ability to aggregate results for overall model assessment, thus preserving both individual and global fairness information
Data Source
Figure 1A
Figure 1B
Figure 2A
AI summary
State of the art model fairness approaches do not address the degree of local fairness of a ML model. A method and system for determining local fairness of a classification Machine Learning (ML) model with degree of fairness is disclosed. The method creates multiple perturb instances using multilevel GMM clustering approach and a constrained perturbation technique to ensure feature distribution of perturbed data, generated from a tabular base data is within the feature distribution of the tabular base data of the ML model. Further, the class of a protected attribute is flipped, black box model prediction probabilities and the cosine similarity constraint and multiplication factor on the probabilities is used to provide a degree of fairness for the local instance. Thus, provides magnitude of fairness or unfairness to the local instance.