Bias evaluation program, apparatus, and method
The bias evaluation device identifies and corrects biases within specific groups by classifying data and calculating a DI score, addressing masked unfairness in machine learning models.
JP7868691B2Active Publication Date: 2026-06-02FUJITSU LTD
Patent Information
- Authority / Receiving Office
- JP Β· JP
- Patent Type
- Patents
- Current Assignee / Owner
- FUJITSU LTD
- Filing Date
- 2022-11-16
- Publication Date
- 2026-06-02
AI Technical Summary
Technical Problem
Existing bias detection methods in machine learning models fail to identify biases within specific groups within the overall data population, masking unfairness that is not apparent when comparing to the entire population.
Method used
A bias evaluation device that classifies data into multiple groups based on attributes, identifies a second group with lower positive occurrence rates, and calculates a DI score by comparing it with other groups to detect and correct biases.
Benefits of technology
The method effectively detects and corrects biases within specific groups that are masked by overall comparisons, ensuring fair training data for machine learning models.
β¦ Generated by Eureka AI based on patent content.
Smart Images
Figure 0007868691000006 
Figure 0007868691000007 
Figure 0007868691000008
Abstract
A bias evaluation device according to the present invention sorts a plurality of pieces of data into a plurality of groups on the basis of a first attribute of a plurality of attributes that each of the plurality of pieces of data has, specifies a second group of the plurality of groups for which the positivity rate of the data included is lower than the positivity rate of the data included in a first group x of the plurality of groups, and performs a bias evaluation on the data on the basis of a comparison of the first group and another group y of the plurality of groups that does not include the second group.
Need to check novelty before this filing date? Find Prior Art