Bias Evaluation Across Subgroups for Hidden Fairness Gaps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models can produce discriminatory results due to biased training data, which is not detected when comparing a group to be evaluated with all other groups, leading to hidden unfairness and potential social loss.
Innovation Solution
A bias evaluation device and method that classifies data into groups based on attributes, identifies subgroups with lower positive example ratios, and calculates a modified DI score to detect and correct biases by comparing the evaluated group with selected subgroups, generating training data to ensure fairness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If data is divided into groups and a ratio of positive example ratios is used as an index to indicate fairness, then fairness evaluation can be performed, but biases hidden in specific subgroups cannot be detected
Solution Approach 1:
The patent segments the evaluation process into multiple stages: first dividing data into groups based on attributes, then further dividing each group into subgroups. This hierarchical segmentation allows detection of biases at different levels, revealing hidden subgroup biases that would be missed by simple group-level comparison alone.
Solution Approach 2:
The patent introduces a new dimension to the evaluation by calculating disparity indices not only at the group level but also at the subgroup level within each group. This multi-dimensional approach transforms a single-level comparison into a hierarchical multi-level analysis, enabling detection of biases that exist only within specific subgroups.
2Reliability
If comparison is made between a group to be evaluated and all other groups, then overall fairness can be assessed, but hidden unfairness in specific subgroups remains undetected
Solution Approach 1:
The patent segments the comparison process by first evaluating groups at a high level, then segmenting each group into subgroups for detailed evaluation. This segmentation prevents information loss by ensuring that subgroup-level biases are captured and reported separately, rather than being masked by aggregate group-level statistics.
Solution Approach 2:
The patent performs excessive evaluation by conducting both group-level and subgroup-level disparity index calculations. This partial or excessive action ensures that no bias information is lost, as the evaluation covers all possible levels of granularity, even though it requires additional computational effort.
3Productivity
If machine learning models use biased training data, then model training efficiency is maintained, but discriminatory results are produced leading to social loss
Solution Approach 1:
The patent applies preliminary action by performing bias evaluation on training data before the machine learning model is trained. The disparity index calculation identifies biased data patterns in advance, allowing data correction or reweighting to occur before training begins. This prevents discriminatory results without requiring retraining, thus maintaining productivity while eliminating harmful factors.
Solution Approach 2:
The patent introduces feedback by using the calculated disparity index to identify and correct biased training data before model training. The evaluation results feed back into the data preparation process, allowing corrective actions to be taken on the training data itself, thereby preventing discriminatory outputs while maintaining training efficiency.
Data Source
AI summary
A bias evaluation device includes a processor that executes a procedure. The procedure includes: classifying a plurality of items of data into a plurality of groups based on a first attribute of a plurality of attributes included in each item of the plurality of items of data; identifying, from among the plurality of groups, a second group having a lower positive example ratio of data included than a positive example ratio of data included in a first group of the plurality of groups; and executing data bias evaluation based on comparison of the first group with another group, not including the second group, among the plurality of groups.