5. Changing the data instead#
An easy way to make an existing cost-insensitive model cost-sensitive is to preprocess the data. Nothing about the estimator changes; the training set it sees does.
The two techniques come from different literatures. Cost-proportionate sampling draws directly from the cost matrix. Bias mitigation takes a page from fairness research and removes the model’s bias against a subgroup, which you can define strategically to serve a business goal — in a churn problem, for instance, customers with a high lifetime value.
How class imbalance relates to cost asymmetry, and where resampling sits next to changing the objective or the threshold.
Resample the training set in proportion to the cost matrix, making any estimator cost-sensitive.
Remove the model’s bias against a subgroup you define — high-value customers, for instance.