Machine Learning Fairness Correction Data Influence Assessment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models trained with biased data can lead to unfair determinations, and fairness correction processes may degrade the accuracy of these models, necessitating a method to assess the influence of training data on model accuracy while maintaining fairness.
Innovation Solution
A determination program and apparatus that process data to identify attributes with significant contribution to inference results, calculating a processing amount and contribution magnitude to determine the influence degree of corrected data on the machine learning model, thereby selecting optimal correction data for retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If fairness correction process is applied to retrain the machine learning model with processed training data, then fairness of determination is improved, but accuracy of the machine learning model deteriorates
Solution Approach 1:
The patent applies partial action by selectively processing only certain training data samples that exhibit unfairness, rather than uniformly processing all data. The determination apparatus identifies specific data points where the machine learning model produces unfair determinations and applies correction processes only to those samples, thereby minimizing the impact on overall model accuracy while still improving fairness.
Solution Approach 2:
The patent changes parameters of the training data by adjusting determination results for specific attributes when unfairness is detected. The system modifies output parameters (determination results) based on detected unfairness patterns, transforming the data in a controlled manner that preserves useful information while eliminating biased outcomes.
2Reliability
If training data is processed to correct fairness issues, then fairness of determination is improved, but the influence of training data on model accuracy increases
Solution Approach 1:
The patent implements a feedback mechanism where the determination apparatus continuously monitors the machine learning model's determination results, detects unfairness patterns, and uses this information to guide subsequent data processing and retraining decisions. This closed-loop feedback allows the system to adjust the degree of data processing based on observed effects, preventing excessive modification that would harm accuracy.
Solution Approach 2:
The patent replaces traditional mechanical data processing methods with an intelligent, adaptive system that uses detection and determination algorithms. Instead of applying fixed preprocessing transformations, the system dynamically identifies unfairness patterns and applies targeted corrections, substituting rigid mechanical processing with flexible intelligent control.
Data Source
AI summary
A determination program for causing a computer to execute processing including: identifying, based on a difference between a first plurality of pieces of data and a second plurality of pieces of data obtained by processing the first plurality of pieces of data based on nonuniformity of the first plurality of pieces of data with reference to a first attribute out of a plurality of attributes, at least one second attribute processed with a processing amount larger than or equal to a predetermined threshold out of the plurality of attributes; identifying a magnitude of contribution of the at least one second attribute to an inference result in a case where data is input and a machine learning model performs inference; and determining, based on the magnitude of the contribution, an influence degree in a case where the machine learning model is trained by using the second plurality of pieces of data.