How to deal with address (like zip-code) for training a model?

To me it doesn't make sense to normalize it even if it is a numerical variable like Zip Code. An address should be interpreted as categorical features like neighborhood... ?

Suppose I have geolocalisation data (latitude longitude), the best thing to do seem to use k-means clustering and then working with cluster's label that I encode.

If the answer is : it depends please tell me how