2 ms·
It is not possible to avoid racial bias. If you make a model predict equally for two classes, you’ve likely only made your predictions worse for both. Predicti
by throwawayuuuu5 7y ago
It is not possible to avoid racial bias. If you make a model predict equally for two classes, you’ve likely only made your predictions worse for both.
Prediction outputs come with a degree of confidence. It is much better to change how you evaluate your predicted probabilities for each protected class than to try and force the probabilities themselves to come out even.
The algorithm is rarely going to be the problem. It is the input data that is allowed into the model. If you have a problem with how healthy whites are getting recommended for treatment more than sick blacks, you need to remove the inputs that otherwise differentiate them.
- LanceH 7y agoOutreach to healthy people is a different care than for someone who is sick. The comparison should be between two groups with the same need/types of care.
- throwawayuuuu5 7y agoThat’s backwards. If the problem is that the algorithm says healthy white gets treatment more than sick black; that means the algorithm has something else that it’s considering besides healthy and sick. If that’s the problem you’re solving for, you need to remove the input that creates the difference. Once it’s gone, the algorithm will be unable to produce a different result for two people with the same sickness level.