Bias Detection in Machine Learning Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models often exhibit biases when trained on certain types of input data, leading to inaccurate performance compared to other inputs, and existing methods fail to effectively detect and address these biases before model generation.
Innovation Solution
A computer-implemented method and system that analyze training datasets to identify and correct label bias by detecting statistically significant differences between sensitive and non-sensitive groups, and generate separate machine learning models for each group to mitigate sampling bias, using techniques like propensity score matching and cluster analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single machine learning model is trained on heterogeneous data with sensitive groups, then model generation is simplified, but bias arises causing inaccurate performance for certain input types
Solution Approach 1:
The patent segments the training data into sensitive and non-sensitive groups based on sensitive parameters, and generates separate machine learning models for each group. This segmentation approach resolves the contradiction by maintaining simplified model generation processes for each homogeneous group while ensuring reliable and accurate performance across all input types through group-specific optimization.
2Reliability
If bias detection and correction is performed before model training, then model accuracy improves, but processing time and computational resources increase
Solution Approach 1:
The patent performs bias detection and correction as preliminary actions before model training by analyzing training data for label bias and sampling bias, and correcting identified biases in advance. This preliminary action resolves the contradiction by ensuring high model accuracy through pre-correction while managing processing time through efficient bias detection algorithms and targeted corrections only where needed.
3Reliability
If separate models are generated for sensitive and non-sensitive groups, then bias is reduced improving model reliability, but device complexity and resource utilization increase
Solution Approach 1:
The system segments models into separate sensitive and non-sensitive group models, managed through an organized architecture that reduces perceived complexity
Solution Approach 2:
The patent introduces an intermediary bias detection and correction system that automates the segmentation and model generation process. This intermediary layer resolves the contradiction by handling the complexity of separate model management automatically, reducing bias through proper segmentation while presenting a simplified interface to users and reducing overall system complexity through automation.
4Measurement precision
If bias correction techniques are applied to training data, then label accuracy improves, but data processing complexity increases
Solution Approach 1:
The patent applies bias correction techniques as preliminary data processing steps before model training, identifying and correcting label bias in the training dataset in advance. This preliminary action resolves the contradiction by ensuring high label accuracy through systematic correction while managing data processing complexity through automated detection algorithms and structured correction procedures that streamline the overall process.
Data Source
AI summary
There is provided computer implemented method for detecting and reducing or removing bias for generating a machine learning model, comprising: prior to generating the machine learning model: receiving a training dataset, comprising target inputs, each comprising parameters and labelled with a corresponding target output, wherein at least one of the parameters of at least of the target inputs comprises a sensitive parameter indicative of the corresponding target input assigned to a sensitive group that is potentially biased against other target inputs that are excluded from the sensitive group, analyzing the training dataset to identify target inputs affected by label bias when a statistically significant difference is detected between target inputs assigned to the sensitive group and target inputs excluded from the sensitive group, correcting labels of the target inputs affected by label bias, and generating the machine learning model using the corrected labels.