Debiasing System for Multi-Choice Recommendation Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing debiasing techniques are not adaptable to address bias reduction in multi-choice models, and they often limit the utility of recommendation models by dropping predictors.
Innovation Solution
A system and method that detect bias and perform debiasing during the model deployment and usage stage, using a debiasing system with modules for determining fairness ratios, sorting scores, thresholding, and adjusting scores to ensure fairness ratios fall within predetermined thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing debiasing techniques are applied to multi-choice models, then bias detection capability is improved, but the techniques are not adaptable and fail to maintain predictive power
Solution Approach 1:
The patent changes the parameter being measured from simple demographic distribution to fairness ratios that compare predicted outcome distributions across different demographic groups. This parameter transformation enables the debiasing system to work effectively with multi-choice models by measuring whether protected groups receive proportional representation in recommendation outcomes, thereby resolving the adaptability issue while maintaining precise bias detection
Solution Approach 2:
The debiasing system is designed to be universally applicable across different model types (binary choice and multi-choice) and different recommendation scenarios. The fairness ratio calculation methodology works regardless of the specific model architecture or recommendation context, making the system adaptable to various multi-choice models while maintaining consistent bias detection capabilities
2Reliability
If existing debiasing techniques are applied, then bias reduction is attempted, but predictive power is lost due to dropping predictors
Solution Approach 1:
Instead of removing predictors from the model, the patent extracts bias from the output predictions by calculating fairness ratios and adjusting predicted outcomes. The original predictors remain intact in the model, preserving predictive power, while the debiasing operation separately modifies the recommendation outputs to ensure fairness across demographic groups
Solution Approach 2:
The patent introduces fairness ratios as an intermediary metric between the model predictions and the final recommendations. This intermediary layer allows the system to maintain the original predictive relationships while mediating the output to ensure fair representation across protected groups, thereby preserving predictive power while achieving fairness
3Reliability
If fairness ratios are adjusted to meet thresholds, then fairness is improved, but model score accuracy may be affected
Solution Approach 1:
The patent applies partial adjustment to model scores, modifying only the portion necessary to achieve fairness thresholds rather than completely recalibrating all scores. This selective adjustment approach ensures fairness compliance while minimizing impact on the overall accuracy and precision of the model scores
Data Source
AI summary
A system and method detect bias and to perform debiasing of an item including data, ideas, and processes, by determining a fairness ratio, from model scores from a recommendation model, sorting the model scores into an ordered list of scores, determining that a sorted score is within a range of predetermined upper and lower thresholds, adjusting the sorted score, recomputing the fairness ratio from the adjusted sorted score, and in the case that the recomputed fairness ratio is not within the range, further adjusting the sorted score until a corresponding fairness ratio of the adjusted sorted score is within the range. In the case that the recomputed fairness ratio is within the range, the system and method generate and output a message with the adjusted sorted score to a recommendation engine, which generates and outputs a recommendation based on the adjusted sorted score.