Hacker News Reader

Comment by tern | Hacker News Reader

Wow! Music recommenders come up every once in awhile on HN and they're almost alway unusable for underground music. Aside from not recognizing some of the names I tried, this one does quite well. It just needs quick links to listen to examples and it'll already be more useful than Spotify, Apple Music or Last.fm.

Some test cases I used: Logic1000, Tessela, Piezo, TSVI, Two Shell, Caterina Barbieri, Schacke, Amnesia Scanner, Doss, Celyn June [fail], DJ Plead [fail], DVRTN [fail], JakoJako [poor results], LDS [needs disambiguation]

This may already be a better approach, but has anyone built a music recommender that's based on published data (stuff you can find on Discogs) like record labels, aliases, collaborators, remixed-by, etc? As a listener mainly to dance, experimental, and contemporary classical music, I suspect a system based on simple metadata associations would easily beat other approaches.

View on HN

That's exactly what I've been working on. Choose it based on the people someone chose to work with and those networks.

It's very much just a single user system that I haven't generalized because that's a way harder problem, the rules aren't generalizable. Classical music for instance, may have 15 names on the credits, a jazz record may have like 4 labels it gets placed on.

Let's take an EP by a popular artist. It may have remixes by popular DJs. Those links are poor quality. Now if it's by an unknown artist, those links usually become high quality.

So do you follow the network of the guy playing the oboe? Maybe? Sometimes weird connections like the album artist is the strong link, sometimes it's compilation albums that one of the songs is placed on. Take say the 1992 release Trancemaster 1: https://www.discogs.com/release/54719-Various-Trancemaster-V... the clustering of those artists is a very strong high quality link. And then there's the "that's what I call music" type compilations where they're worthless.

I can do the music I like because I can narrow the ruleset but a general application is basically a winograd schema challenge because there's a large body of intuition required to weight the network. This task is certainly a nontrivial neural network problem.

Doing it manually with human discretion works. I've got tools for doing that and large labeled data sets I've been working on for 4 years. It just doesn't generalize.

Some day ...

Nice! Glad you're doing this work. Sounds like something that would be best integrated into a community with strong contributor culture that could do the tagging, but that's another task entirely.

That makes sense. I threw in my favorite (used to be underground) artist and saw every side-project they had pop up (and they have around a dozen). OTOH, I just threw in Nujabes and saw a number of people I do not recognize, which I find interesting.

From my tests, the suggestions make sense. It seems to be better than the spotify "I will play things that send you into an echo chamber" algorithm.

There's two approaches: musicologist and popularity.

The musicologists do a serious study and have sophisticated tools I can't pretend to understand but their results sound similar: so similar that the serendipity and adventure is sucked out of it.

Also they suffer from the generality problem as well. Qualities that matter in one genre, such as airy female vocals in a minor key, or whatever, are absolutely irrelevant for another genre. Take for instance, Irish folk music where it may be signal and say, Acapella, where it may be noise. Thus considering them is both right and wrong and we're back to our problem.

If you want to categorize the music to "fix" the problem, you run into binning issues. Let's say early 90s hardcore; they can have trance, breakbeat, dnb, house and jazz sections in a single song. Good luck trying to use your genre based contextual mapping on an unsupervised model.

The popularity approach, which ignores the content, is deceptive because it appears in many forms: people who listened to X also listened to Y or Y is trending or any of a number of variations where some magnitude of humans or temporal delta is used as signal.

These all tend towards the not long end of the long tail and so you eventually get the same mediocre experience - you start with your obscure prog rock group from the 1960s and 20 songs later you're at Cream or Hendrix. The "solution" is to tamper the drift via clustering but it will tend the same directions.

In practice though, these approaches service the majority of tastes, that is familiarity, so they're fit for purpose.

Hilariously King Stingray maps next to Wet Leg ...

I'm guessing that's because they're both recent breakout bands based on an island . . .

KS - https://www.youtube.com/watch?v=sr3iI8gg2fo

WL - https://www.youtube.com/watch?v=Zd9jeJk2UHQ

That would have to be a specialized genre tracker though, most of your suggestions would fall flat for metal, and even label would not always be helpful.

FWIW, the site does not work for my underground bands at all, out of 5 tries 3 didn’t show up at all (Remember Twilight, Der Rest, Jessica's Crime), and 1 (Scythia) has barely any related bands. Only "Die Streuner" got decent results. Some of the bands it’s missing even have wikipedia pages.

I also looked into more obsure bands I'm into. I suspect the relatedness is affected by the band lineups of previous tours

haha nice tested with similar artists (eartheater, doon kanda, dean blunt...) :)) works really well. i'm impressed!

dean <3

could this be as simple as...a python script which takes an artist - using python discogs_client and just scours every label mate of that artist and feeds itt to you - made more intelligent with acoustic fingerprinting...and a tiny bit of human intervention? :)

For certain scenes and genres, I suspect this would already be very useful

Reply on news.ycombinator.com