I have found this to be terrifically difficult. I went to the trouble of going through Tom Waits' discography to find something similar to "Alice," without luc

I have found this to be terrifically difficult. I went to the trouble of going through Tom Waits' discography to find something similar to "Alice," without luck. Some of my best mixes take years to create. The problem, to me, breaks down into three separate difficulties, perhaps three and a half: First, if the Music Genome people are correct, the number of variables (if you will) is quite large, creating an enormous search space. I grant an extra half to the concept that portions of that search space would be sparsely populated. We are getting pretty good at searching these kinds of spaces (something I would like to know more about). Second, each person has a different idea of what similarity entails. For some it is instrumentation. Others, production value. Some are looking for lyrical subject similarity. And so forth. And most people are very much "I'll know it when I hear it," unable to really qualify that similarity function. It requires quite a bit of introspection, meta-cognition, and a good vocabulary to really describe what that target is. Right now, we're bad at that and have a bad habit of just throwing everything in a bucket. As an example, I might like some music with a fast tempo for some mood, but I might also like "darkjazz" (a microgenre which more or less collapsed about seven or so years back). Most recommender systems would then pick out some fast, "hot" jazz, which I absolutely do not want, but it has the tempo of some things I do like and the genre of something else I like. Third, the process of assigning values to said variables was quite labor intensive when they were doing it, as well as subjective. Highly trained musicians spending half an hour on each song (which means that, with that many variables, while the time seems long, the examination must be cursory). It would really be something to make progress in this area and I think it might be the most fruitful. However, I think any real progress in this area might mean you would need a corpus of at least a hundred thousand songs human-rated, each by multiple evaluators, then programs written for each of the, say 450 genes (if we are to use the Music Genome Project). Finally, the programs would be tweaked until they aligned with the ratings given by humans. It's one of those things I would fund if I were an eccentric billionaire.