Besides ease of implementation, due to the certainty of having binary splits, what are the advantages of coding categorical features into dummy variables in the context of decision trees?
Does using dummy variables speed up the tree growing process? Does it reduce underfitting/overfitting?
For a categorical variable with many levels, representing it as a set of dummys might not be the best way! since it forces splits only on one given level. There might well be better splits into two sets of levels, but searching among all such splittings is prohibitive. Many other ideas are possible, and some can be found from here: Random Forest Regression with sparse data in Python
