Matrix factorization can deal with this a bit by using the high dimensional space to place your tastes into an area that reflects many different styles at the same time. In part 2 of the blog post we're going to talk about how we're modeling the acoustic qualities of music, which can find common patterns from completely different genres (for example, you may like soothing music with female vocals in both jazz and indie rock). In part 3 we'll talk a bit about how we can combine recent signals (like thumbs) to take into account your current mood, which I find helps pinpoint interesting music to surface right now.
That's neat! I was curious of this is / was being looked into. It seems like I often get music that's matched based on a demographic (if that makes sense), rather than music matched on the characteristic features of the current song / band.
The worst case of that is probably "new age", a label rejected by virtually all the artists so labeled (and most of the listeners), and having no common traits to speak of, but lumped together as whatever sold better in bookstores than in record stores.
This is the most interesting part of the problem to me. I always worry that the signals I send my favorite music radio service permanently alter the course of the channel. In some cases, I want that. In others I don't.
They're mapping the multidimensional music spectrum. I highly doubt the end goal/use is to firmly place you in some 10d music coordinate system so you can listen there forever. You need the mapping before you can do anything interesting, like a cool auto generated real life soundtrack that uses data from your pulse, phone calls, paycheck, work hours, interactions, etc. :D
I think clustering in vector space works really well for many things, but not for discovery of new unexpected music.
The only thing in my experience that works for that is algorithms that take advantage of human curation (people with similar tastes). And even then filter bubble is a real thing (Facebook)
After you downvote "Faithfully" enough times because it makes you gag, Journey leaves your list.
Except that "Frontiers" and "Edge of the Blade" sound nothing like the garbagey Journey ballad schmaltz that everybody sings on karaoke night. So, you will never hear those songs in spite of the fact that you may like them.
The problem is that you can't get this information without throwing the occasional curveball at a listener. And, from the streaming app point of view, it is very risky to throw new songs at a user that you don't know definitively that they like because they might dislike it enough to change the channel, app, etc.
Consequently, recommendation engines run by corporations will only ever be totally safe and boring.
This is the problem with machine learning or most algorithmic recommendation schemes. There are no curveballs, no randomness. Of course you want one part of the experience to be similarity.
I really want the equivalent of visiting a friend's house and them putting on a mixtape and a couple of the tracks or artists suddenly jumping out at you. Or going to a gig and being blown away by a support band. Or a dinner conversation about a new band your friend just discovered, or ...
That probably won't ever happen - I can't see an algorithm recommending Julie London if you usually listen to Moby, but some of those bizarre leaps are often the best discoveries.
On your first trip to the record store, you won't have any idea what you're buying. It might be good, it might not be good. After a while, you get to know the owner of the store, and they learn your tastes, and you learn a bit more about music and start to know what will be good and what won't be good. Sometimes you'll both make a mistake and you'll still buy something you don't like. And you'll learn from that.
Why shouldn't our learning algorithms work the same way?