ML Discriminatory Region Rules for Fair Classification Screening

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Machine-learning models can introduce discriminatory bias, classifying samples differently based on protected attributes like age, ethnicity, or gender, which is unacceptable in certain applications, and traditional techniques struggle to effectively identify and mitigate this bias.

Innovation Solution

A central server analyzes a machine-learning model's output classifications using a training set of samples with known protected attribute values to generate rules delineating discriminatory regions, allowing the model to be used while preventing discriminatory classifications by rejecting samples that fall within these regions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a machine-learning model is trained to classify samples efficiently, then classification productivity is improved, but discriminatory bias may be introduced that classifies samples differently based on protected attributes

Engineering Contradiction:

Improveclassification efficiencyVSAvoiddiscriminatory bias

Core Design Contradiction:

ProductivityVSObject-affected harmful factors

Solution Approach 1:

The patent segments the classification process into two distinct stages: (1) a machine-learning model performs efficient initial classification of samples, and (2) a separate rule-based system evaluates the output classifications to identify and correct discriminatory bias. This segmentation allows each component to specialize - the ML model optimizes for efficiency while the rule-based system optimizes for fairness - resolving the contradiction between productivity and harmful discriminatory effects

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediary rule-based system that acts as a mediator between the machine-learning model and the final classification output. This intermediary evaluates the ML model's predictions against predefined fairness rules related to protected attributes, and can override discriminatory classifications. The intermediary preserves the efficiency benefits of ML classification while eliminating harmful bias, resolving the technical contradiction

Inventive Principle:

Principle #24Intermediary (Mediator)

2Difficulty of detecting and measuring

If traditional techniques are used to identify discriminatory bias, then detection capability is limited, but system complexity is kept low

Engineering Contradiction:

Improvediscriminatory bias detection capabilityVSAvoidsystem complexity

Core Design Contradiction:

Difficulty of detecting and measuringVSDevice complexity

Solution Approach 1:

The patent replaces traditional manual or simple statistical techniques for detecting discriminatory bias with an automated rule-based evaluation system. This system programmatically assesses ML model outputs against predefined fairness criteria related to protected attributes, significantly enhancing detection capability. The rule-based approach automates what would otherwise require complex manual analysis, improving detection while maintaining manageable system complexity through structured, interpretable rules

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Object-affected harmful factors

If the machine-learning model is retrained to eliminate discriminatory bias, then fairness is improved, but training time and computational resources are consumed

Engineering Contradiction:

Improvediscriminatory biasVSAvoidmodel retraining time

Core Design Contradiction:

Object-affected harmful factorsVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-defining fairness rules and evaluation criteria before the machine-learning model is deployed for classification. Instead of retraining the model when bias is detected, the pre-established rule-based system can immediately evaluate and correct discriminatory outcomes. This preliminary preparation eliminates the need for time-consuming retraining cycles, as the fairness enforcement mechanism is already in place and can operate independently of model retraining

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent extracts the fairness enforcement function from the machine-learning model itself and places it in a separate rule-based evaluation system. This extraction allows the ML model to focus on its strength (efficient classification) while the extracted fairness component handles bias mitigation. By separating these functions, the system avoids the need to retrain the complex ML model for fairness improvements - instead, only the simpler rule-based system needs to be updated, significantly reducing time and computational resource requirements

Inventive Principle:

Principle #2Taking out (Extraction)

Data Source

PatentUS12555002B2Rule generation for machine-learning model discriminatory regions

Publication Date: 2026.02.17 INTERNATIONAL BUSINESS MACHINE CORPORATION

AI summary

One embodiment provides a method, including: providing, from the central server to a machine-learning model, a training set of samples having known values for at least one target protected attribute, wherein the training set includes a first set of samples having a first value for the at least one target protected attribute and a second set of samples having a second value for the at least one target protected attribute; receiving, at the central server from the machine-learning model, an output classification for each of the samples within the training set of samples; and generating, at the central server using the output classification, a set of rules delineating a region within the machine-learning model as discriminatory, wherein the region includes a classification region where the machine-learning model classifies received samples differently based upon a value of the at least one protected attribute.