Machine Learning Module Bias Control Through Feature Correlation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning modules often exhibit bias towards specific subsets of input features, leading to unfair predictions or control decisions.
Innovation Solution
A method for generating a machine learning module that includes training data correlation metrics, bias metric calculation, and subset definition to ensure the module is fair by satisfying a bias constraint, using automated processes to adjust model parameters and feature subsets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning modules are trained on existing data without bias control, then training speed and simplicity are improved, but prediction fairness and reliability deteriorate due to bias towards specific feature subsets
Solution Approach 1:
The system performs preliminary actions by calculating correlation metrics between features and target variables before training begins. It identifies and flags potentially biased feature subsets in advance, allowing the training process to account for these biases through appropriate weighting or adjustment mechanisms, thus preventing unfair predictions while maintaining training efficiency
Solution Approach 2:
The system implements feedback by continuously monitoring prediction outcomes across different feature subsets during and after training. When bias is detected in certain subsets, the system adjusts training weights or model parameters to compensate, ensuring that the final model produces fair predictions across all demographic or feature groups while maintaining overall training productivity
2Reliability
If bias control mechanisms are implemented in machine learning modules, then prediction fairness is improved, but system complexity increases due to additional metrics and constraints
Solution Approach 1:
The system segments the feature space into distinct subsets based on demographic or sensitive attributes. By analyzing and controlling bias for each segment separately rather than treating all data uniformly, the system can implement targeted fairness constraints that are more manageable and less complex than global fairness mechanisms, while still achieving overall prediction fairness
Solution Approach 2:
The system changes parameters by introducing correlation metrics and bias constraints as additional mathematical parameters in the training objective function. These parameters are integrated into the existing optimization framework, allowing bias control to be achieved through parameter adjustment rather than fundamentally redesigning the system architecture, thus limiting the increase in complexity
3Measurement precision
If correlation metrics and bias constraints are calculated during training, then prediction accuracy and fairness are improved, but computational time and resources increase
Solution Approach 1:
The system applies partial action by calculating correlation metrics and bias constraints only for critical feature subsets that have the greatest impact on prediction fairness, rather than comprehensively analyzing all possible feature combinations. This selective approach maintains prediction accuracy for the most important groups while significantly reducing computational time and resources compared to exhaustive bias analysis
Data Source
AI summary
The present disclosure relates to a computer-implemented method for generating a machine learning module (ML-module). The method comprises providing training data comprising a first set of data corresponding to a first feature and a second set of data corresponding to a second feature and generating a further set of data, wherein the further set of data corresponds to a further feature. A first correlation metric as a measure of a correlation between a selected feature of the first and the second feature and the further feature is calculated. Furthermore, a bias metric is determined indicating a strength of a bias of the trained ML-module towards a first subset of values of the further feature or a second subset of values of the further feature if the first correlation metric is greater than a first threshold. If the bias metric satisfies the bias constraint, then the ML-module is released for usage.