This changes the outcome psuedoCount parameter from a scalar to a map. This lets us input our prior beliefs about the outcome class distribution.
A limitation of the current approach of using a single scalar for pseudocount is that it is only applied to outcomes that have been previously observed in the training data.
When the model hasn't received a lot of training data it will predict with high confidence the outcomes it has already seen, regardless of pseudocount.
It is common to know which classes exist even though they haven't been observed so it is useful to be able to use that knowledge for predictions.
This changes the outcome psuedoCount parameter from a scalar to a map. This lets us input our prior beliefs about the outcome class distribution. A limitation of the current approach of using a single scalar for pseudocount is that it is only applied to outcomes that have been previously observed in the training data. When the model hasn't received a lot of training data it will predict with high confidence the outcomes it has already seen, regardless of pseudocount.
It is common to know which classes exist even though they haven't been observed so it is useful to be able to use that knowledge for predictions.