AI Bias Detection via Perturbed Transaction Records
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI models used for determining outcomes can exhibit bias, which is challenging to detect efficiently without significantly impacting the resources required for testing various inputs, especially in batch scoring scenarios where large datasets are processed.
Innovation Solution
A method and system that create perturbed transaction records by altering attributes of original records, allowing for the analysis of bias in AI models while reducing the number of records sent for processing, using a bias detection model to identify and score bias in outcomes, and storing these scores in a database for further analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If a subset of original transaction records is selected to reduce the number of records sent to the outcome model, then resource consumption is reduced, but the accuracy of bias detection may be compromised
Solution Approach 1:
The system performs preliminary actions by creating perturbed transaction records before sending them to the outcome model. For each selected original transaction record, multiple perturbed versions are generated with modified attributes, and these perturbed records are processed in advance to detect bias patterns. This preliminary processing of a reduced subset enables efficient bias detection while maintaining accuracy through the perturbation technique.
2Measurement precision
If perturbed transaction records are created by changing attributes of original records, then bias detection capability is improved, but the complexity of the processing system increases
Solution Approach 1:
The system creates copies of original transaction records and modifies their attributes to generate perturbed versions. Instead of developing complex new detection mechanisms, the invention copies existing records and systematically alters their attributes (such as demographic characteristics) to test for bias. This copying approach simplifies the overall system architecture while enabling comprehensive bias detection through attribute modification.
3Loss of time
If a subset selection criteria is applied to reduce the number of records processed, then processing time is reduced, but the representativeness of the sample may be reduced
Solution Approach 1:
The system applies parameter changes by systematically modifying attributes of selected transaction records to create perturbed versions. The subset selection criteria focuses on diverse records that represent different demographic groups and outcome scenarios. By changing parameters such as demographic attributes in the perturbed records and comparing outcomes, the system maintains sample representativeness while processing a reduced subset, ensuring reliable bias detection across different population segments.
Data Source
AI summary
A system and related method score a fairness of an outcome model. The method comprises receiving a set of original transaction records (OTRs), and selecting an OTR subset of the OTRs according to a subset selection criteria in order to reduce a number of OTRs to send to outcome model. For each OTR in the subset a perturbed transaction record (PTR) is created based on the OTR that includes changing at least one attribute in the PTR from the OTR, sending the OTR and the PTR to the outcome model, receiving an OTR outcome and a PTR outcome from the outcome model, and determining a record bias score for the OTR outcome and the PTR outcome respectively that indicates bias in the respective outcome. The OTR and the PTR bias score are stored in a bias determination system (BDS) database.