Seems like a logical approach to me. Also once you get the data set built you could algorithmically generate new recipes. If one type of machine learning algorithm generates more pleasing new cocktails you have a winner. Just don't do all the testing yourself. We don't want to read the HN post explaining how 'sparkpluglabs' ended up in detox from unit testing!