Predictive Model Bias Correction via Class Normalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing predictive models used in digital interviewing and other evaluation platforms often incorporate adverse impacts, particularly biases against protected classes, which are difficult to detect and correct, leading to unfair outcomes in hiring, education, and other fields.
Innovation Solution
A model training tool that processes digital interview data to identify and correct adverse impacts by using a normalization process, such as a weighting factor algorithm, to ensure each candidate class has equal weight, and incorporates a penalty function to adhere to legal thresholds, effectively managing unstructured data types like audio and video.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional predictive models are trained on historical performance data, then the model can learn important features for prediction, but adverse impacts and biases against protected classes are incorporated into the model
Solution Approach 1:
The training data is segmented by protected class (e.g., minority vs. majority groups) to identify and isolate adverse impacts. The model training process is divided into multiple stages: initial training on all data, then separate training on segmented data to detect biases, and finally correction by adjusting weights or removing biased features for each segment.
Solution Approach 2:
The patent modifies training parameters by introducing weighting factors that adjust the influence of different protected classes during model training. By changing the weight parameters assigned to each class's training data, the system can counterbalance historical biases and ensure fairer prediction outcomes across all groups.
2Reliability
If the model is trained to improve prediction accuracy, then it may exploit biased patterns in the training data, but this reduces fairness and increases adverse impact
Solution Approach 1:
The system implements feedback loops where prediction outcomes are continuously monitored for adverse impact across protected classes. When bias is detected in the model's predictions, the system automatically adjusts training weights or removes problematic features, creating a closed-loop system that maintains both reliability and fairness through iterative correction.
Solution Approach 2:
The patent applies preliminary anti-action by proactively identifying and correcting biases during the model training phase before deployment. Rather than waiting for unfair outcomes to manifest, the system preemptively adjusts training parameters and removes biased patterns, preventing adverse impact from occurring in the first place.
3Ease of operation
If normalization processes are applied to correct adverse impact, then fairness improves, but model complexity and processing requirements increase
Solution Approach 1:
The normalization process applies local quality by targeting only specific features or data segments that exhibit adverse impact, rather than uniformly processing all data. This selective approach corrects biases in affected areas while leaving the rest of the model architecture intact, thereby improving fairness without proportionally increasing overall system complexity.
Data Source
AI summary
Technology for training a predictive model is described. A processing device collects digital interview data including features identified from candidate interviews. A model training tool selects a data set of the digital interview data. The data set includes a predicted performance outcome and an actual performance outcome for each of a plurality of candidates. The model training tool determines an error metric for each of the plurality of candidates. The error metric includes a relationship between the predicted performance outcome and the actual performance outcome for each candidate. The model training tool determines a number of candidates whose digital interview data includes a feature corresponding to a protected class. The model training tool normalizes an effect of each candidate on the error metric based on the corresponding protected class and applies the normalized error metric to reduce bias in the predictive model with respect to the protected class.