Mitigating Governance Bias in Machine Learning Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models are susceptible to governance-induced bias due to data governance policies and privacy laws that restrict the use of personal and sensitive data, leading to biased training datasets without mechanisms to detect or mitigate these biases effectively.
Innovation Solution
A method that involves obtaining a governance impact assessment for a partial data set, identifying core features affected by data governance, and applying anti-bias procedures to mitigate the introduced bias, using metadata to log and analyze data exclusion and obfuscation decisions, and employing feature relevance functions to assess the impact of excluded data on model performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data governance policies are applied to restrict personal data usage, then data privacy and compliance are improved, but training data quality and model accuracy deteriorate due to excluded features
Solution Approach 1:
The system performs preliminary actions by generating governance impact assessments before model training to identify which features will be excluded due to data governance policies. This allows proactive mitigation strategies to be applied, such as alternative feature engineering or data synthesis, to compensate for the excluded features and maintain model accuracy while ensuring compliance
Solution Approach 2:
The governance impact assessment serves as an intermediary mechanism that bridges data governance requirements and machine learning model training. It provides detailed information about excluded features and their potential impact, enabling the system to mediate between compliance constraints and model performance requirements through informed decision-making
2Ease of operation
If comprehensive data governance enforcement is applied, then data subject control and transparency are improved, but bias in machine learning models increases due to excluded data
Solution Approach 1:
The system implements feedback mechanisms by continuously monitoring governance impact assessments to identify features that are systematically excluded and may introduce bias. The assessment provides feedback about the nature and extent of exclusions, enabling iterative adjustment of mitigation strategies to maintain fairness while respecting data governance requirements
Solution Approach 2:
The system applies parameter changes by modifying model training parameters, feature selections, or data preprocessing steps based on governance impact assessment findings. This allows the system to adapt to the constraints imposed by data governance policies while compensating for excluded features through parameter adjustments that maintain model fairness and reduce bias
3Object-affected harmful factors
If data governance policies exclude certain features, then data privacy protection is improved, but feature completeness and model performance deteriorate
Solution Approach 1:
The system applies copying by creating alternative representations or synthetic versions of excluded features that do not contain sensitive personal information. The governance impact assessment identifies which features are excluded and enables the creation of proxy features or synthesized data that preserve the predictive value of excluded features without violating privacy requirements
Data Source
AI summary
A method, a computerized apparatus and a computer program product for mitigating governance and regulation implications on machine learning. A governance impact assessment is generated for a partial data set generated by applying a data governance enforcement on a data set of instances comprising valuations of a feature vector. The partial data set comprises partial instances each comprising partial feature vectors. The governance impact assessment comprises information about data excluded from the data set. A machine learning model trained based on the partial data set and configured to provide an estimated prediction for a partial instance is obtained. A set of core features is determined. A bias introduced by the data governance is identified based on a core feature being affected by the data governance. In response to identifying a bias, an anti-bias procedure is applied on the machine learning model, whereby mitigating the bias introduced by the data governance.