GAN Dataset Generation for Algorithmic Fairness Evaluation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for evaluating fairness in Machine Learning (ML) models lack consistent and systematic evaluation practices, rely on small and unrealistic datasets, and fail to account for diverse bias patterns, leading to inconsistent and non-reproducible results.

Innovation Solution

A method and system for generating anonymized datasets with predetermined biases using Generative Adversarial Networks (GANs) to create realistic and diverse datasets for evaluating algorithmic fairness, allowing for fine-grained control over bias types and scenarios, thereby enhancing the robustness of fairness evaluations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If small existing datasets (e.g., COMPAS, UCI Adult) are used for fairness evaluation, then the evaluation can be performed with available data, but the datasets are unrealistic and lead to inconsistent and non-reproducible results

Engineering Contradiction:
Improvereproducibility of fairness evaluationVSAvoiddataset size
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent uses Generative Adversarial Networks (GANs) to create synthetic copies of real-world datasets that preserve the statistical properties and bias patterns of original data while removing sensitive information. This allows reproduction of realistic fairness evaluation scenarios without relying on small, inconsistent existing datasets

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent systematically varies parameters such as bias strength, group size ratios, and feature distributions in generated datasets to create multiple versions with controlled characteristics. This enables consistent reproduction of different fairness challenge scenarios while maintaining realistic data properties

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If diverse bias patterns are introduced to evaluate model robustness, then the evaluation becomes more comprehensive, but the complexity of dataset generation and evaluation increases

Engineering Contradiction:
Improvecoverage of bias patternsVSAvoidevaluation system complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent segments bias introduction into distinct, modular components: group size disparity, prevalence disparity, and distinct conditional class separability. Each bias type can be independently controlled and evaluated, making the complex evaluation process more manageable and systematic

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The GAN-based dataset generation framework serves multiple functions: it can generate datasets with different bias patterns, remove sensitive information for privacy protection, and maintain realistic statistical properties. This multi-functional approach reduces overall system complexity despite the comprehensive evaluation capabilities

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Reliability

If GANs are used to generate anonymized datasets, then privacy is protected and large realistic datasets are created, but the generation process is computationally intensive and time-consuming

Engineering Contradiction:
Improveprivacy protectionVSAvoiddataset generation time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent performs preliminary actions by pre-training GANs on real-world data to learn underlying distributions before actual fairness evaluation. Once trained, the GANs can rapidly generate synthetic datasets with controlled bias patterns, reducing generation time for subsequent evaluations

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent optimizes GAN generation parameters such as batch size, number of epochs, and randomness seeds to balance privacy protection with generation speed. By carefully tuning these parameters, the system achieves both strong anonymization and reasonable generation times

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4141693A1Method and device for obtaining a generated dataset with a predetermined bias for evaluating algorithmic fairness of a machine learning model
Publication Date: 2023.03.01 FEEDZAI CONSULTADORIA E INOVACAO TECHCA SA
  • EP4141693A1 patent drawingFigure 1~2
  • EP4141693A1 patent drawingFigure 3~4
  • EP4141693A1 patent drawingFigure 5~6A

AI summary

Computer system and method for obtaining a generated dataset with a predetermined bias, from a real-world tabular dataset, for evaluating algorithmic fairness of a machine learning model, comprising training a Generative Adversarial Network, GAN, model on the real-world dataset for generating a dataset as a anonymized reconstructed dataset of the real-world dataset, and sampling of the trained GAN model to obtain the generated dataset. Introducing the predetermined bias in the generated dataset may comprise appending one or more columns generated to introduce bias into the dataset; selecting one or more columns and sampling the generated dataset according to a predetermined distribution of the selected column or columns for introducing bias into the dataset; or by adapting a value function of the GAN for introducing the bias in the training of the GAN model on the real-world dataset.