Bias Detection in Machine Learning via Consistency Evaluation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models often suffer from bias, particularly labeling bias, which is difficult to address due to human factors such as background and views, leading to inconsistent and inaccurate training data.
Innovation Solution
A method that evaluates labeled training data to detect anomalies, cross-validates these anomalies by prompting other humans for labeling, differentiates valid labels from biased ones, and augments training data to reduce bias by collecting more representative samples, considering the profiles of human labelers to provide accurate classifications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If human labeling is used to create training data, then the training data can be generated with domain expertise, but human background and views introduce labeling bias and inconsistency
Solution Approach 1:
The patent introduces an intermediary consistency evaluation mechanism that mediates between multiple human labelers. This intermediary system evaluates labeling consistency across different annotators and identifies biased labels, thereby resolving the contradiction between utilizing human expertise and eliminating human-induced bias.
Solution Approach 2:
The patent implements a feedback loop where labeling results are evaluated for consistency, and this evaluation feedback is used to identify and correct biased labels. The system continuously refines labeling quality by feeding back consistency metrics to the labeling process, improving both accuracy and reliability iteratively.
2Reliability
If multiple human labelers are used to evaluate labeled data, then bias detection capability is improved, but the complexity and time consumption of the process increases
Solution Approach 1:
The patent segments the bias detection process into distinct modules: initial labeling, consistency evaluation, anomaly detection, and correction. By dividing the complex process into manageable segments, the system achieves high reliability in bias detection while keeping each individual step relatively simple and controllable.
3Reliability
If labeled output data with low consistency assessment is selected for additional labeling, then bias reduction is improved, but the time and resources required for re-labeling increase
Solution Approach 1:
The patent applies partial action by selectively re-labeling only those data instances that fall below the consistency threshold, rather than re-labeling the entire dataset. This targeted approach achieves effective bias reduction while minimizing the time and resources required, as only the problematic portion of the data receives additional labeling attention.
Data Source
AI summary
A method identifies and removes bias from a machine learning model. A user/computer inputs a plurality of input training data into a machine learning system to generate an output of labeled output data. The user/computer evaluates the labeled output data according to a consistency metric to associate the labeled output data with a corresponding consistency assessment. The user/computer selects each labeled output data having a consistency assessment indicating a consistency assessment that is greater than a predetermined threshold to form a labeled output data subset, and then creates additional labeling for the labeled output data subset. The user/computer utilizes the additional labeling to distinguish each labeled training data from labeled output data subset as being mislabeled and biased, and then adjusts the learning machine based on the labeled output data subset being mislabeled and biased.