Adaptive Fairness Repair Pipeline for Unlabeled Bias Mitigation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional machine learning bias mitigation strategies are limited by restrictive data requirements, lack generalizability across different models, and fail to address multiple biasing attributes, often introducing new biases and requiring labeled data, which is expensive and hard to acquire.
Innovation Solution
An iterative fairness learning process that applies adaptive fairness refinement techniques to reduce bias across multiple attributes using unlabeled data, adjusting model outputs iteratively to minimize divergence between probability distributions, applicable to both classification and regression models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional bias mitigation strategies are used, then bias reduction is achieved for single attributes, but data requirements become restrictive and labeled data is required
Solution Approach 1:
The patent changes the fundamental parameter of data usage from requiring labeled data to utilizing unlabeled data. The fairness learning model trains on unlabeled data by learning transformations that equalize probability distributions across different attribute groups, eliminating the need for expensive labeled data while maintaining bias mitigation effectiveness
Solution Approach 2:
The patent creates a universal bias mitigation approach that works across multiple attributes simultaneously. The single fairness learning model can handle multiple biasing attributes (e.g., race, gender, age) together, making the system universally applicable to various fairness scenarios without requiring separate models for each attribute
2Reliability
If traditional bias mitigation strategies are used, then bias is addressed for one attribute at a time, but generalizability across different models is limited
Solution Approach 1:
The fairness learning model is designed to be universally applicable across different machine learning models and attributes. It learns a general transformation that can be applied to multiple attributes simultaneously, making the system adaptable to various fairness scenarios without requiring model-specific or attribute-specific separate models
Solution Approach 2:
The patent segments the bias mitigation process into two independent stages: a fairness learning stage that learns transformations from unlabeled data, and a fairness application stage that applies these transformations to labeled data. This segmentation allows the system to be trained on one type of data and applied to another, enhancing generalizability
3Reliability
If multiple biasing attributes are addressed simultaneously, then comprehensive fairness is improved, but computational resources and complexity increase
Solution Approach 1:
The patent performs preliminary learning of fairness transformations during an unsupervised stage on unlabeled data before applying them to the actual bias mitigation task. This preliminary action separates the complex learning process from the application process, allowing multiple attributes to be handled efficiently by pre-computing the transformation rules
4Measurement precision
If labeled data is used for training, then model accuracy is improved, but acquisition cost and time increase
Solution Approach 1:
The patent fundamentally changes the data requirement parameter from labeled to unlabeled. The fairness learning model learns transformations by analyzing probability distributions across attribute groups in unlabeled data, eliminating the need for time-consuming data labeling while still achieving accurate bias mitigation
Solution Approach 2:
The patent creates a copy of the labeling process through unsupervised learning. Instead of requiring actual labeled data, the system learns the essential patterns and transformations by analyzing the structure of unlabeled data, effectively copying the informative content without needing the expensive labels
Data Source
AI summary
Various embodiments of the present disclosure provide improved bias mitigation techniques for machine learning technology. The bias mitigation techniques include receiving a machine learning biasing attribute and a set of outputs generated by a target machine learned model and determining a divergence score based at least in part on a difference between a first distribution and a second distribution of the set of outputs. In response to the divergence score meeting or exceeding a threshold, the bias mitigation techniques include generating, using a fairness learning model, a transformation to one or more set of outputs to decrease the divergence score and storing the transformation and an identifier. Through a plurality of iterations, a learned transformation sequence may be generated that may be applied to transform outputs of a machine learning model to mitigate machine learning bias.