ML Bias Detection Using Class-Aware Model Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models, particularly those performing regression, suffer from bias that is not adequately addressed by conventional methods, which are often focused on binomial classification and lack generalizability, necessitating a more effective approach to detect and reduce bias in continuous output models.
Innovation Solution
A method involving training a baseline model with a classification algorithm to generate a class-aware model, comparing outputs, and iteratively removing correlated parameters to correct bias, using a bias detection engine to ensure fairness across classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional bias reduction methods are used, then binomial classification bias is addressed, but continuous output regression model bias remains unaddressed
Solution Approach 1:
The patent creates a universal bias detection and reduction framework that works across both binomial classification and continuous output regression models. The system uses a unified approach of training multiple models (baseline, class-aware, and threshold-aware variants) and comparing their outputs to detect bias, making the methodology applicable to diverse model types rather than being limited to specific classification scenarios.
Solution Approach 2:
The patent systematically varies key parameters including threshold values, model configurations, and training data subsets to detect how model outputs change under different conditions. By analyzing output variations across multiple parameter settings, the system identifies biased parameters and reduces their impact, providing a generalizable approach that adapts to different model types and contexts.
2Reliability
If manual bias reduction processes are used, then bias can be addressed, but the process becomes tedious and requires skilled intervention
Solution Approach 1:
The patent implements an automated bias detection and reduction system that performs multiple model trainings, output comparisons, and parameter identifications without human intervention. The system automatically generates confidence intervals, identifies biased parameters based on output discrepancies, and produces corrected models, eliminating the need for manual analysis and reducing the process from a tedious skilled task to an automated self-service operation.
Solution Approach 2:
The system employs feedback loops where model outputs are continuously compared across different configurations, and the results feed into automatic parameter identification and model correction. This iterative feedback process automatically refines the models and reduces bias without requiring manual review at each step, significantly reducing time loss while maintaining high detection accuracy.
3Measurement precision
If baseline model training is performed, then a starting point for bias detection is established, but class-specific biases remain undetected
Solution Approach 1:
The patent segments the model evaluation process by creating multiple specialized models: a baseline model for overall performance, class-aware models for specific class behavior, and threshold-aware variants for different decision thresholds. This segmentation allows bias detection at multiple granularities, identifying both general and class-specific biases that would be invisible in a single monolithic model evaluation.
Solution Approach 2:
The patent adds dimensional complexity to the evaluation by training models with different awareness dimensions (class-awareness, threshold-awareness) and comparing outputs across these dimensions. This multi-dimensional approach reveals biases that manifest differently across various model configurations, providing comprehensive detection capability despite the increased number of models required.
Data Source
AI summary
Techniques are described herein for verifying and/or reducing bias in machine learning (ML) models. An ML model may be trained using a baseline training data set (TDS). A second TDS may be generated from the baseline TDS based on executing a classification algorithm to assign each example to one of a set of classes. The class-aware ML models may be trained using class-specific examples (e.g., examples that correspond to a given class). Bias in the ML model may be detected based on providing an input to the ML model and each of the class-aware ML models and identifying differences between their respective outputs. If bias is detected, a parameter that highly correlates to the output of the ML model may be removed from the training data and the process may begin anew. Each iteration of the process may reduce the bias originally detected in the ML model.