Blogs/Addressing Bias and Fairness via Separate Classifiers for Each Group

Addressing Bias and Fairness via Separate Classifiers for Each Group

peterwashington Nov 01 2021 2 min read 0 views
Bias and Fairness
BiasFairnessSeparateClassifiers.png

Even with oversampling of the data from underrepresented groups, the model may still not learn as well for this group due to the limited distinct examples it sees. After balancing the dataset, a further technique can be to train a classifier for each of the individual subgroups and combine the output of the relevant subgroup classifier with the classifier trained on all of the data.

For example, say we have a dataset which predicts the likelihood of a pediatric disease, and most of the data in our dataset are for very young children (ages 0 to 9). In addition to building a classifier for the entire dataset, we could build a classifier trained only on a small age group. To make a prediction, we could combine the outputs for the overall classifier with the age-specific classifier:

In the above example, the overall classifier predicts a probability of 0.54 for having the disease for the test datapoint where the child’s age is 11, but the age group-specific classifier predicts a probability of 0.97. To get a final prediction, we could average these two output probabilities to get a final probability of 0.755. While the overall classifier barely predicted that the child had the disease, the age-specific classifier seemed to output a much stronger prediction.