Prototype-Based Debiasing With Adaptive Sample Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models trained on biased datasets, particularly those with underrepresented groups, exhibit poor performance and unintended discrimination, as they are difficult to identify and control biases in.
Innovation Solution
A prototype-based learning method that iteratively adds prototypes to underrepresented classes, re-trains the model, and computes sample weights based on class-wise quality metrics to ensure unbiased training and evaluation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning models are trained on available patient data, then the model can be trained and deployed, but the model performs unintentionally different on certain subgroups leading to discrimination
Solution Approach 1:
The patent segments the training data analysis by creating prototypes for different subgroups (e.g., racial groups, geographic regions) and calculating separate quality metrics for each prototype. This segmentation allows the system to identify and address biases in specific subgroups without retraining the entire model on all data, thus maintaining training efficiency while improving fairness.
Solution Approach 2:
The patent performs preliminary analysis by calculating quality metrics and identifying biased subgroups before final model deployment or before problematic biases manifest. By pre-identifying underrepresented groups and calculating their quality metrics, the system can proactively adjust training strategies or data sampling to prevent discriminatory outcomes while maintaining efficient training workflows.
2Reliability
If weighted sampling is used to increase representation of underrepresented groups, then bias in the model can be reduced, but it is difficult to choose the appropriate sampling weights
Solution Approach 1:
The patent implements a self-service mechanism where the system automatically calculates quality metrics for different subgroups and determines appropriate sampling weights without requiring manual intervention. The system serves itself by identifying biased subgroups, computing their quality metrics, and adjusting data sampling accordingly, thus reducing the complexity of weight determination while maintaining model fairness.
Solution Approach 2:
The patent dynamically adjusts sampling parameters based on calculated quality metrics. By changing the sampling weights as a parameter based on the measured quality of different subgroups, the system automatically optimizes representation without requiring complex manual weight determination. The parameter adjustment is driven by empirical quality measurements rather than theoretical assumptions.
3Reliability
If prototypes are added to underrepresented classes and the model is retrained, then dataset biases are automatically identified and addressed, but additional computation and training time are required
Solution Approach 1:
The patent applies partial action by adding prototypes only to underrepresented classes where quality metrics indicate bias, rather than retraining the entire model uniformly. This selective approach addresses biases in specific subgroups without performing exhaustive retraining on all data, thus reducing overall training time while maintaining accurate bias identification.
Solution Approach 2:
The patent implements a dynamic process where prototypes are iteratively added based on quality metric thresholds. The system dynamically adjusts the number and distribution of prototypes across classes, retraining only when and where needed. This dynamic adaptation allows the system to achieve reliable bias identification while minimizing unnecessary retraining time through conditional, targeted updates.
Data Source
AI summary
A prototype-based model with one prototype per class is analyzed based on a comparison between a class-wise quality metric and a threshold for each prototype per class pair. A new prototype is added for each class in response to the class-wise quality metric for a respective prototype per class pair being below the threshold. The prototype-based model is retrained using the new prototype. A number of training samples closest to each prototype per class pair of the retrained prototype-based model is counted based on a distance measure. Sample weights for the training samples closest to each prototype per class pair is computed based on the number. A target model is trained using the computed sample weights. The method has applications including, but not limited to, use cases in computational biology, medical AI and healthcare, and cyber threat security for optimizing machine learning processes or supporting decision making.