Post-Training Model Bias Mitigation

Now that we understand the risks of models and how damaging model bias can be, let’s identify some theoretical solutions. Approaches in this space come in two flavors: post-training and in-training.

Post-training methods are a way to normalize the behavior of the algorithm after training is done. These methods are typically not as powerful as in-training ones, but are a lot faster and easier to implement.

Randomization

Randomization is one of the simplest and most effective (but worst-performing) methods of debiasing an algorithm. The premise of this is to reduce the accuracy on the majority group to equalize performance with the minority group.

If a skin cancer detection algorithm has a 90% accuracy on the majority group and an 80% accuracy on the minority group, the simplest fix would be to randomize with some probability whenever the majority group is encountered. For example, if the majority class is the $X$ variable, randomize the assigned class 20% of the time.

Mathematically, this equalizes the accuracies quite well. The minority accuracy stays the same, but the majority probability is the original accuracy (90%) with probability 0.80, and a randomized accuracy (50%) with probability 0.20.