I have been looking through questions regarding categorical feature encoding, but couldn't find any which discuss my problem. Apologies if I missed it.
Let's say we have a dataset with binary and nominal variables of roughly equal importance each.
Most classifiers cannot deal with categorical types directly, so these have to be transformed - for example using one-hot encoding (dummy variables) as explained in this answer.
If one categorical variable has high cardinality, wouldn't encoding it this way "overpower" other (for example binary) variables? By "cardinality" I mean the number of categories in a nominal variable.
If our classifier model is aware of relationships between variables, wouldn't it unnecessarily attempt to find relationships between introduced binary dummy "components" of the same variable?
And if so, how could this be addressed?
The best solution I can think of is to logically group high-cardinality properties into "buckets", however if there are enough unique values to be a problem, then manually grouping them would be labour consuming as well.
Edit: This is trivial and only partially addresses the problem, but one of the things I ended up doing is replacing all relatively rare categorical values with a new, "other" category. It could be time consuming to optimise the threshold when to consider value "rare", but at least this approach can be automated.