Active Learning Bias Reduction via Data Point Modification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Active learning in machine learning lacks a systematic approach to reduce bias in training data and models, leading to potential unfairness in AI decision-making.
Innovation Solution
A method that monitors data and model bias during the training process, modifies data points within a predefined tolerance to ensure they fall within a desired class, and discards points that do not meet these criteria, thereby reducing bias and enhancing fairness in the training dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If active learning is used to train machine learning models with fewer training labels, then training efficiency is improved, but bias in the training data and model increases
Solution Approach 1:
The system continuously monitors bias metrics during the active learning training process. When bias exceeds predefined thresholds, the system provides feedback by modifying the active learning query selection to prioritize data points that reduce bias, thereby maintaining training efficiency while correcting bias accumulation in real-time
Solution Approach 2:
The system dynamically adjusts the bias threshold parameters and query selection criteria based on the current state of the training dataset. By changing these parameters adaptively, the system maintains optimal training efficiency while preventing bias from exceeding acceptable levels
2Reliability
If data points are modified to reduce bias, then model fairness is improved, but data accuracy may deteriorate
Solution Approach 1:
The system applies bias reduction modifications selectively and locally to specific data points rather than uniformly to all data. By identifying and modifying only those data points that contribute to bias while preserving accurate data points, the system improves fairness without sacrificing overall data accuracy
Solution Approach 2:
The system applies partial modifications to data points, adjusting them just enough to reduce bias while staying within predefined tolerance limits. This partial action approach ensures fairness improvement without excessive modification that would compromise data accuracy
3Measurement precision
If bias monitoring is performed at each addition of data to virtual training dataset, then bias detection precision is improved, but computational complexity increases
Solution Approach 1:
The system performs bias monitoring at each data addition but applies partial checking strategies, focusing computational resources on monitoring critical bias metrics and data points that have the greatest impact on model fairness, thereby maintaining high detection precision with reduced computational overhead
Solution Approach 2:
The bias monitoring process is segmented into modular components that can be independently executed. By dividing the monitoring task into separate bias metric calculations and threshold comparisons, the system achieves precise bias detection while reducing overall computational complexity through modular processing
Data Source
AI summary
Aspects of the present invention provide an approach for reducing bias in active learning. In an embodiment, a data point is selected from a training dataset for a current training iteration while monitoring for data bias at each addition of data to a virtual training dataset. In addition, a machine learning model is examined for bias after adding the selected data point to the virtual training dataset. When data bias and/or model bias is detected, the data point is considered for potential label modification. The selected data point is modified and, if the raw value of the modified data point is within a predefined tolerance and within a bin of a desired class, the modified data point having a label of the target class is retained. Otherwise, it can be discarded.