Closed dscolby closed 3 months ago
We should also include a note in the documentation for this.
Since an ELM is minimizing an MSE loss, we can do this for binary outcomes by clipping values outside of [0, 1] but this isn't really feasible for categorical variables. For example, if there are four categories, 1, 2,3, and 4, the model predicts 1.6, and the actual class is 4, you can't interpret this as having a higher probability of being 3 than 4 because these are just arbitrary categories. Therefore, we will only support clipping functions for binary treatments and outcomes.
Closing since this isn't feasible.
This is essentially the same thing as https://github.com/dscolby/CausalELM.jl/issues/36. We will use the clipping function because it preserves class predictions and constrains the predictions within the range of natural values.