Causal Bayesian Network Bias Reduction in ML Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models often suffer from algorithmic bias, particularly when using non-protected attributes as proxies for protected attributes, leading to discrimination and reduced predictive power, as conventional approaches fail to adequately address causation-based discrimination.
Innovation Solution
The implementation of a causal Bayesian network approach for supervised machine learning, which models the entire process to reduce or eliminate confounder bias by training with observed data and applying default values or omitting protected attributes during inference, allowing for the use of new data sources while minimizing bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine learning approaches use non-protected attributes as proxies for protected attributes, then the model can capture predictive patterns, but algorithmic bias and discrimination occur
Solution Approach 1:
The patent segments the feature set into protected attributes and non-protected attributes, allowing the model to process them differently. During training, both types of features are used to maximize predictive power. During inference, protected attributes are omitted or replaced with default values, eliminating bias while preserving the model's ability to learn from non-protected features that capture genuine predictive patterns.
Solution Approach 2:
The patent performs preliminary action by training the model with both protected and non-protected attributes during the training phase, allowing the model to learn optimal feature representations. Before inference, protected attributes are pre-processed by omitting them or replacing with default values, so the model applies learned patterns without biased inputs during actual prediction.
2Object-affected harmful factors
If protected attributes are omitted from training data, then bias is reduced, but predictive power decreases due to loss of information
Solution Approach 1:
The patent segments the feature processing into training and inference phases. During training, protected attributes are included alongside non-protected attributes, allowing the model to learn from complete information. During inference, only non-protected attributes are used, preventing bias while the model leverages patterns learned from both feature types during training.
Solution Approach 2:
The model performs preliminary learning during training with full feature sets including protected attributes. This preliminary action allows the model to internalize predictive patterns that may be captured through correlations with protected attributes. During inference, the model applies these learned patterns using only non-protected attributes, maintaining predictive power without bias.
3Object-affected harmful factors
If a causal Bayesian network approach is implemented to model the entire process, then fairness and bias reduction are achieved, but device complexity increases
Solution Approach 1:
The patent extracts and removes protected attributes from the inference process while retaining them during training. This extraction eliminates the source of bias (protected attributes) from the decision-making process while preserving the causal relationships and predictive patterns learned during training, achieving fairness without requiring complex debiasing mechanisms.
Solution Approach 2:
The causal Bayesian network performs preliminary analysis during training to identify causal relationships between features and target variables. By understanding the causal structure in advance, the model can make fair predictions during inference using only non-protected attributes, as the causal relationships have already been encoded in the model parameters during training with complete data.
Data Source
AI summary
One or more default protected attribute values may be determined for a prediction model trained based on training data including a plurality of training observations. Each of the plurality of training observations may include a respective plurality of training data values corresponding with a plurality of features. Each of the plurality of training observations may also include a respective target value. Each of the plurality of training observations may include a respective protected attribute value corresponding with a protected attribute feature. A request to determine a designated predicted target value for a designated inference observation may be received after determining the one or more default protected attribute values. The predicted target value may be determined by applying the prediction model to an inference observation and a designated default protected attribute value of the one or more default protected attribute values.