[sparkpluglabs]

May be I am not thinking right about it.

Let me give an example. If given a list of all cocktails, I have lot of positive examples on what ingredients go well together. Say Gin and Vermouth. So I have big list of positive examples. But is hard to get negative examples.

I want to build a model where given some ingredients the model tells me whether they go well together or not.

[mswen]

What I would do is build a crowd-sourced data set of what ingredients go well together and which ingredients people don't like together (rate it from 1 to 10, where 1 means these ingredients should never be used together to 10 means this is a heavenly combination). After lots of rating by many different people you could have a average quantitative score for each pair of ingredients.

You now have variation and can use a variety of predictive approaches

[sparkpluglabs]

Thanks!. Your comment gave another idea. What if I take a cocktail recipe as a vote. I take all ingredients and create generate all possible pairs. Look for how many times the pair appear in cocktails. Treat those as votes. More votes means they go together really well.

[mswen]

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!

[sparkpluglabs]

:) a "surprise me" button for http://www.kickstarter.com/projects/monsieur/monsieur-the-ar...