Human-in-the-Loop Fairness Optimization in Machine Learning with Minimax Loss and an Abstain Option
Description
Statistics
Description
Input: List of features corresponding to an individual. More information can be found in the Features tab.
Output:
Architecture:
Prediction confidence:Probability score of a prediction
Dataset
Choose one of the following datasets
Given a set of features for an individual, predict whether that individual makes more than $50k per year
Given a set of features for an individual, predict whether that individual is going to recividate
Algorithm
Loads a test sample and makes a prediction
Reset changes
Similar Samples
Settings
Cost
Predictions where the algorithm is less confident can have manual review or help; however, having human experts review these predictions is costly for time and resources.
This slider controls how likely the algorithm will require manual review. The higher the cost, the more likely the system is to say "I don't know" and request human intervention.
Current Value:
Accuracy-Fairness
Controls the accuracy - fairness tradeoff. Highest means the algorithm favors fairness significantly more than accuracy. Fairness refers to the discrepancy between the prediction's confidence level and the original data (e.g. it is most unfair when it's 99% confident about a wrong prediction for an individual sample). Accuracy is defined as the percentage of correct predictions.
Current Value: