Band level doesn't quite work. Alas, it needs to be track level. There are a lot of bands where I like a track and that's it.

My favorite example is Seven Sirens And a Silver Tear from Sirenia, a Norwegian gothic metal band. There's no metal in that. It took me a long long time before I learned this but the track is a direct descendant of the Midlight Sonata. And I was hunting for similar songs and I now keep a playlist of them -- but if you started from Sirenia you would never find any of them.

This. There are many, many artists who I only own one or two tracks of - including some of the most-played tracks in my collection - because the vast majority of their other output is not my taste at all. If finding good music was as simple as just buying everything that a single artist put out, it would be much easier to build a collection.

The good news is that in the digital era you no longer need to fork over cash for an entire album or even an EP when you only care about one of the songs - which leaves more money available to buy music from other artists. I often wonder if in the long run it still balances out for artists, since the songs one person likes probably aren't the same as the songs another person likes, especially in niche genres.

If the track is known, then from the context of this and other things you like, you would still reach useful recommendations in not too many clicks. Track level would be interesting, but is also harder as data is a lot sparser making it harder to build reliable recommendations.

What is the approach, on a concrete, technical level, that you are taking to make recommendation N, based on the 3, 4...N-1 choices?

Do you think online NNMF collaborative filtering with Spotify bands with fewer than 100,000 monthly listeners is the answer? If you had infinite resources, what would you do?

This is incredible!

I wish I could give it more bands, and see the distance (I imagine it computes one?) between each band I provided and the ones it suggests.

Did you install malware on my computer ;) How did it manage to predict so much of what in my collection with just 3 names ?

I tried it before, and I just tried it again. It is great. I just came back from RecSys a large conference on recommender systems. Researchers and companies like Spotify, Amazon music, Deezer gave lots of presentation. However, nothing they showed were so immediately useful as this. Awesome service!

Some interesting suggestions. However, from a number of starting points I was pointed as "Songraes", a band that doesn't seem to rate it's own Wikipedia article. I wonder if this has been "fixed" somehow?

Quick update: Gnoosic has helped me find a bunch of great bands. Thank you! Definitely a great tool :-)

Neat, thank you! I'll definitely give it a shot and see how it goes.

Wow. Simple user interface, fas and it gives interesting results! It did not find two of my favorite groups, gusgus and subgud, but I added a suggestion. Bookmarking this for later use!