I like to implement something I have already done before. In my case, the ID3 algorithm has a nice balance of challenge, experience and documentation available. You could try to write it for a specific case, where you structure your data, and then apply it to a generic case.
