I'd really like to know where this data comes from. The results are accurate enough, but how can I know they'll become more accurate or improve over time?
To incentivize people to keep contributing, the plan is to build a notification service around your favorite artists, so you'll find out about new artists/albums/songs that are related to your tastes.
So the more popular the site gets, the more data there will be :)
IMHO, it doesn't seem correct to ask people for their "favorite bands". It would seem more accurate to ask them "when you're in a given mood, what bands will you most enjoy hearing?". At the right times, I'd be equally happy to hear Symphony X or Johnny Cash. But at any given time, I'm going to prefer one or the other. So why not just let the contributors spell out those similarity sets for you?
It seems to me that you're grouping things that are similar in some stronger sense than that they're the favorites of some people. And conversely, the way you're getting the data -- asking for favorites you'd be missing out on a bunch of minor bands that aren't anyone's favorite, but still enjoyed by people.
Last.fm does alright: http://www.last.fm/music/Dark+Moor/+similar
Last.fm puts Galneryus on page 3, and Galneryus is sufficiently different that I wouldn't trust the recommendations that score lower. Last.fm correctly places Stratovarius above this cutoff, but Helloween, Sonata Arctica, and Blind Guardian don't make the cut, and Ayreon is all the way down on page 10 with Van Canto and Turisas. Van Canto and Turisas are great, to be sure, they are just way less similar to Dark Moor than Ayreon is.
Of course, Last.fm is better than nothing: http://ifyoudig.net/mohican-sandbag vs. http://www.last.fm/music/Mohican+Sandbag/+similar
Last.fm just reads tags, so a lot of artists have multiple names. モヒカンサンドバッグ is the second result for being similar to itself! Last.fm also does a pretty awful job by putting ORANGE★JAM and 3L on the first page, with dBu closely following. Much more similar producers and circles like 和泉幸奇, Alstroemeria, or IOSYS are nowhere to be found. Izmizm and Shibayan on page 1 are sensible. So maybe Last.fm isn't all that much better than nothing...
I think this can be good, and I might even pay for it, but ultimately it will be just one of a handful of half-solutions that need to be combined to get quality discovery. Pandora and Google Music's instant playlist do better than I would expect any people-who-like-x-also-like-y similarity system to do. Unfortunately Pandora has like 7 artists in its library and Google Music's instant playlist requires you to already have the music you are discovering, which sort of misses the point.
Any algorithms anyone can suggest for finding close spellings, besides Levenshtein? Like, that are somehow indexable or easy to implement in a database?
[Edit: just updated it, searching now works on a direct string match, no more annoying alert box.]
Jokes aside, having something show up on the page would probably work better.
I searched "Don Caballero" on ifyoudig [1] and got a few good results, but also some very different artists, like Mogwai and Aphex Twin.
I think Mogwai and Aphex Twin are pretty good recommendations for someone who likes the math rock/post rock that Don Cab play.
ifyoudig.net seems to have a very small popularity bias compared to e.g. Spotify though, so that's nice.
These services are clever, but I usually know all of their suggestions. :-(
And yes, I'm trying to do better.
Services that rely on bigger services like Spotify (not sure where this site gets its music, that's just an example) create a bit of a barrier to entry for exactly the type of artist would would benefit most from a bit of extra exposure.
It seems that many new "discover new music" services I see these days have exactly this problem.
What's not so good is if you are a genre expert it doesn't do too deep of a dive.
For instance I love the roots artist Guy Clark but I know every artist that is recommended (http://ifyoudig.net/guy-clark). I like the majority of them so the suggestions aren't off by any means but it would be nice to get more obscure (but still accurate) suggestions.
PS I also second the other comments that the type ahead could be a little more lenient.
Some suggestions: this pure social graph approach could be vastly improved for music recommendation by aggregating tags, à la last.fm, or adding music-related features yielded by some waveform analysis.
For instance, I typed in Tame Impala and got these results in this order: Real Estate - Girls - Beach Fossils - Toro Y Moi - Washed Out - Wavves - James Blake. The first three relate well to modern psych rock of Tame Impala, but then things get a little strange: two chillwave acts, one correctly similar psych/noise rock act and a dubstep/downtempo artist!
It's not really a criticism, looking up stuff from the days I listened to bands rather than songs it seems pretty accurate, but I dunno how I'd use it in the present
Right now I can see that liking Artist A makes me 10x more likely to like Artist B. But Artist B might still be really terrible, so it might work out to only a 1% chance I'll actually like Artist B.
Remarkable recommendations. My favorite artist is http://ifyoudig.net/mike-oldfield and actually ifyoudig.net recommended some really interesting bands.
Way the go!
Sometimes it works much better than stuff you get with bootstrap/UIkit/Foundation/whatever.
Need any help on the front-end side of things?
