Interactive Subgroup Discovery for Bias-Sensitive ML Model Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models often fail to identify meaningful subgroups in big data due to insensitivity to sampling bias, amplifying inherent data biases and lacking sensitivity in data classification.
Innovation Solution
An interactive subgroup discovery method that incorporates human intelligence into data discovery mechanisms, using a mixed-initiative approach combining data-driven and knowledge-driven methods to identify and refine subgroups, allowing users to explore subgroup characteristics and adjust prediction thresholds for improved performance and fairness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are applied to big data, then data classification and subgroup identification can be achieved, but the models fail to identify meaningful subgroups due to insensitivity to sampling bias
Solution Approach 1:
The system implements feedback by computing performance measures for each subgroup and comparing them to identify outlying subgroups. This feedback loop allows the system to detect bias in identified subgroups and refine the analysis accordingly, addressing the insensitivity problem through iterative improvement based on performance measurement feedback.
Solution Approach 2:
The patent segments the population into multiple subgroups based on covariate data and then analyzes each subgroup separately. By partitioning the data and evaluating performance measures for each segment, the system can identify outlying subgroups that may exhibit bias, thereby improving both subgroup identification accuracy and bias sensitivity.
2Productivity
If machine learning models identify subgroups in big data, then data analysis capability is enhanced, but the subgroups identified may only amplify inherent bias of the underlying data
Solution Approach 1:
The system uses feedback through performance measure computation and comparison to detect subgroups that may be amplifying bias. By continuously evaluating subgroup performance and identifying outliers, the system can flag potentially biased subgroups for further review, preventing bias amplification while maintaining high data analysis productivity.
Solution Approach 2:
Performance measures act as an intermediary between the machine learning model and the final subgroup identification. This intermediary layer provides an additional evaluation step that can detect bias amplification before final subgroup selection, allowing the system to maintain productivity while filtering out biased results.
3Measurement precision
If performance measures are computed for each subgroup, then subgroup comparison and identification of outlying subgroups is enabled, but computational complexity increases
Solution Approach 1:
The system segments the computation by calculating performance measures for each subgroup separately rather than analyzing all data together. This segmentation enables precise performance evaluation for each subgroup while managing computational complexity through modular processing of divided data segments.
Data Source
AI summary
Obtain covariates and an outcome data for a population. Partition the population into a plurality of subgroups. Produce outcomes predictions by applying a machine learning model to the covariate data for the population. Establish performance measures based on the outcomes predictions. Compare the performance measures for at least one subgroup to the performance measures for at least one other subgroup. Identify an outlying subgroup for which the machine learning model produces performance measures that are different than the performance measures for one or more other subgroups. Optionally, retrain the machine learning model on additional covariate and outcomes data for the outlying subgroup.