People make playlists for a variety of reasons, according to their specific tastes. Songs that appear together on a functional playlist (workout jams, study music, etc) likely won't have appeared together for the same (or even similar reasons). Contrast that even further against playlists built for aesthetic reasons ('sad songs', 'psychedelic-sounding', 'minimal', etc) and it seems hard to imagine how mining just playlists can lead to good music selections.
Is this algorithm attempting to track context or merely playlist contents?
And for the record, if there is one area that I really want machine learning and "AI" to fall flat on its face with, it's music. It's a shame technophiles want to strip the people out of art and its consumption.
I suppose it is kind of like Pandora. We are actually mimicking word vectors (ala word2vec) but using playlists as the data set. So, in theory, it IS learning the context of songs to make some inferences about the songs themselves.
I absolutely agree that "AI" can't replace knowledgeable humans building thoughtful playlists. I started this project because I can't find new music quickly enough, and listening to any computer generated playlist/radio station gets stale fast. I also rarely find artists that I like more than a handful of their songs, so a lot of artist-centric recommendations miss the mark for me.
In a perfect world, we would feed in reams of carefully curated playlists, learn about different contexts songs appear in, and use some signals from the user to find the types of songs they want to listen to. Two songs don't make for a very strong signal - we have discussed trying to use the user's listening history to help refine the playlist.
> And for the record, if there is one area that I really want machine learning and "AI" to fall flat on its face with, it's music. It's a shame technophiles want to strip the people out of art and its consumption.
Already advocating affirmative action for people in arts? That must mean a lot for "deep art" researchers.
I simply want "deep art" to fail. Big difference. A world dominated by AI-generated art will coalesce us all into one or a handful of dull, lowest-common-denominator aesthetics and rob the world of aesthetic vitality. Not to mention all the artists, many of whom are already scraping for work, that get shut out as peoples' artistic pursuits are gratified instantly.
Apparently it has only trained on 30k songs thus far. This would explain the bias towards pop, as 30k is far too small of a training set to accurately capture the long tail of preferences.
I'm not always getting the greatest results. Mixing Portishead with Perturbator gave me sensical results, but Perturbator with Dance With the Dead should have given me other synthwave/outrun style faux-80s music. Instead it gave me The Notorious BIG, Sound Garden, and the Foo Fighters.
Huh! I also asked for something weird (Frank Zappa + Animal Collective) and got Foo Fighters. I wonder if there's some vector math going on that makes Foo Fighters crop up if the input songs are too orthogonal.
You are probably correct about the vector math! We learn low level embeddings from playlists we scraped from Spotify (with little discretion) and are still working on the algorithm.
We are essentially trying to draw a line between the embeddings of the two songs and find "close" songs to points along that line. This should give us a somewhat smooth transition - at least that's the hope.
I suspect some of it may be that we are using a euclidean line through the vector space, but using cosine distance for similarity. We're still trying to get the hang of using the vectors to build a smooth transition between songs.
We are also tuning our model and training variables, as well as pulling in more playlists, which should help (I hope).
I don't think that's what's going on. Perturbator and Dance with the Dead are basically interchangeable since synthwave music is basically just a bunch of contemporary John Carpenter knock-offs.
We used thousands of Spotify playlists to train on, but obviously more popular genres are going to be heavily represented. For rarer songs, it may not pick up on the association between two songs.
We are still trying to tune our heuristics and scrape more playlists, which should help it to learn rarer songs better.
Hmm, got a lot of very poor recommendations when trying various realistic combinations, the recommendation engine seems to promote pop music very heavily and doesn’t seem to consider lesser played tracks fairly in its analysis. Both Apple Music and pandora do a pretty good job of this by comparison.
This is true. I wish we had their entire playlist library!
This was trained on playlists we scraped from Spotify, so by definition pop music is going to be heavily represented. We are still playing with the training parameters and algorithm for building the playlists.
We also have some ideas on including "rare" or new tracks, but haven't gotten to that part yet.
No idea where the feedback loop is for machine learning, but the playlists are off by miles for anything other than popular genres of pop, (alt-)rock, and some others. No Mozart or Beethoven; Couldn't match Arch Enemy on anything heavy.
We use a shallow neural network to learn about songs from playlists we scraped from Spotify, but to actually build the playlist we use some straightforward vector math.
Rarer songs/genres do tend to be all over the map, and certain super popular songs seem to show up all the time. We are still messing with the training, pulling in more playlists, and working on the playlist algorithm. Every training run produces better results!
From the comments, it does appear that we set the threshold for how frequently a song must appear to be included a little low based on the sparsity of our training data.
We basically try to draw a line between the two songs you chose and find songs close to that line. It can produce some weird results, and it very seldom seems to be the smooth transition we were shooting for.
We are still playing around with the algorithm for generating the playlist, as well as getting more/better training data.
People make playlists for a variety of reasons, according to their specific tastes. Songs that appear together on a functional playlist (workout jams, study music, etc) likely won't have appeared together for the same (or even similar reasons). Contrast that even further against playlists built for aesthetic reasons ('sad songs', 'psychedelic-sounding', 'minimal', etc) and it seems hard to imagine how mining just playlists can lead to good music selections.
Is this algorithm attempting to track context or merely playlist contents?
And for the record, if there is one area that I really want machine learning and "AI" to fall flat on its face with, it's music. It's a shame technophiles want to strip the people out of art and its consumption.