Predictive Model Bias Detection via Protected Data Mediation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Predictive performance models often generate biased outcomes despite excluding protected data, as unethical or illegal biases can emerge during machine processing, particularly in neural networks or decision trees, leading to undesirable differences in user classifications.
Innovation Solution
A system and process that utilize a hardware processor to generate and refine artificial intelligence neural networks by analyzing input data, determining predictive biases, scoring nodes, and excluding biased compartments to produce unbiased predictive models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If protected data is excluded from input to predictive models, then legal and ethical compliance is improved, but bias detection capability deteriorates because unprotected data may still contain biased combinations that cannot be detected without protected data context
Solution Approach 1:
The patent introduces protected data as an intermediary tool that is not used for making predictions but specifically for detecting bias in the model. The protected data acts as a mediator to identify biased variable combinations in the unprotected data, allowing the system to maintain compliance while achieving bias detection through a specialized analysis pathway.
Solution Approach 2:
The patent segments the data usage into two distinct functions: protected data is separated and used exclusively for bias detection purposes, while unprotected data is used for predictive modeling. This segmentation allows each data type to serve its specific purpose without contamination, enabling both compliance and bias detection.
2Manufacturing precision
If initial inputs are individually unbiased, then data quality is improved, but bias introduction occurs during machine processing through variable combinations in neural networks or decision trees
Solution Approach 1:
The patent implements a feedback mechanism where the model's outputs are analyzed to detect bias, and this bias information is fed back to identify and remove biased variable combinations from the model structure. This continuous feedback loop prevents bias generation during processing by iteratively refining the model to eliminate biased interactions.
Solution Approach 2:
The patent performs preliminary bias detection by analyzing variable combinations before finalizing the model structure. By identifying and removing biased combinations in advance, the system prevents bias from being introduced during the machine processing stage, even when individual inputs are unbiased.
3Measurement precision
If complex neural networks are used to improve predictive accuracy, then model performance is improved, but bias complexity increases making it harder to detect and remove biased nodes
Solution Approach 1:
The patent extracts and isolates biased nodes and variable combinations from the complex neural network structure. By identifying specific nodes that contribute to bias and removing or neutralizing them, the system maintains the overall complex structure needed for high predictive accuracy while eliminating the biased components.
Solution Approach 2:
The patent applies local quality control by treating different nodes in the neural network differently based on their bias contribution. Rather than simplifying the entire model, it maintains high complexity where needed for accuracy while locally modifying or removing only the biased nodes, preserving overall model performance while reducing bias.
Data Source
AI summary
A hardware processor can receive sets of input data describing assets associated with an entity. The hardware processor can receive inputs responsive to queries of a user. The hardware processor can individually generate predictive models based on a respective set of input data. The hardware processor can calculate predicted outcomes for the user by applying each of models to the inputs. The hardware processor can generate a user interface comprising the predictive outcomes for the user for each of the predictive models.