The LEAST likely button seems quite useful. Search for "Britney Spears", hit "Show LEAST likely". And you get a list of quality artists from a wide variety of genres.
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.
I wonder where they could have gotten their data from. Their 'about' page says, "If You Dig's results are calculated from the music preferences of a whole lotta real people. Like, a lot."
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 :)
Oh I know. :) That part's only semi-automated, though, I bring the new data in, in manual batches. If the overall artist popularities in the new data don't match the existing data to within a certain degree (and also watching IP addresses and submission time patterns), then that data won't get counted.
Just picking some random favorites, I looked at "Symphony X", and then flipped over to "Least Likely". There are several artists there -- Adele, The Offspring, Johnny Cash -- that I quite enjoy.
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.
This site finds similar artists to the one you search for. It has no reason to expect you'll like The Offspring when you entered Symphony X.
It's like telling a friend you like Band X, and they say "Oh man, if you like Band X, you'll probably like Similar Band Y! or Similar Band Z!" It would be odd if you said you liked Jay-Z and someone recommended The Clash because they're so different, even though it's not uncommon to like both.
As the Contribute question stands, I might put Symphony X and Johnny Cash on my list of favorites, which would make the site more likely to recommend one in response to the other. But in real life, if someone asked me what I'd recommend to a Johnny Cash fan, there's no way I'd answer Symphony X.
And I propose that the way to address this is to not ask the user "what are your [unrelated] favorites", but rather, "pick a recent moment, and tell us what music you would have most enjoyed hearing at that time". That implies a stronger link than simply that the listed artists share a spot on my list of unrelated favorites.
Thanks, you're right. That is now #1 on my to-do list -- maybe just go straight to the artist page if there's an exact string match, or else try to suggest a list of closely-spelled artists?
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.]
This is really cool. I'd love to see it calculate not just likelihoods but actual probabilities.
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.
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.
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.
Subscribing for Google Play Music All Access (yeah, say that seven times fast) opens up the instant playlists ("radio stations") to include all of the music in Google's catalog. But I have found the selection of similar music to be pretty conservative; despite Google's large catalog, the same artists/songs come up every time I play a given radio.
I think the major problem with services like this is that they have things at the far ends of the popularity bell curve fall off. Jcore being one of them. Kind of a chicken and the egg problem where people who like any of a number of niche generas come to the site, find it useless to them, and don't contribute. perhaps some incentive for adding new artists could break this cycle? or even just a link in the "Not Found" page tot he contribution suggesting if they like that band to leave their favorites for others.
This is pretty cool. I'm impressed with the quality of the recommendations and really like how simple the site is. Some info on how you calculate the recs would be interesting. (at least for this thread). I only wish the layout was a little better on my phone. (Sent from iPhone)
PS I also second the other comments that the type ahead could be a little more lenient.
Great stuff! I happen to really like all recommendations that came in for me.
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!
Looking at "contribute", it doesn't really reflect my listening patterns since I starting listening to music almost exclusively by streaming. Once upon a time I bought albums and had favourite artists, but now, for me, it's all about the song. I don't really have a favourite band anymore, and the stuff I listen to most is one or two songs by this act, one or two songs by that act
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
This and other music recommendation services tend to have a problem where if you try a popular artist, it will only return popular artists. I'd really like a service that you input popular artists and then it returns more obscure artists with similar influences.
This is awesome. Typed in one of my favorite bands, found new bands I actually like so far. When I do that on Spotify, I get a bunch of crap that every service counts as "similar", but I hate, or bands I've already heard of a million times.
As an independent artist, I don't see any info on how you might get your own music into the site.
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.
http://ifyoudig.net/page/about
-----
Note for the nerds: the math is legit. We take sample size into account when calculating artist associations, and the likelihood factor is the lowest bound of a confidence interval. Everything is statistically significant. Results are real, yo.
----
48 comments
[ 3.2 ms ] story [ 151 ms ] threadWhat'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.
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.
It's like telling a friend you like Band X, and they say "Oh man, if you like Band X, you'll probably like Similar Band Y! or Similar Band Z!" It would be odd if you said you liked Jay-Z and someone recommended The Clash because they're so different, even though it's not uncommon to like both.
As the Contribute question stands, I might put Symphony X and Johnny Cash on my list of favorites, which would make the site more likely to recommend one in response to the other. But in real life, if someone asked me what I'd recommend to a Johnny Cash fan, there's no way I'd answer Symphony X.
And I propose that the way to address this is to not ask the user "what are your [unrelated] favorites", but rather, "pick a recent moment, and tell us what music you would have most enjoyed hearing at that time". That implies a stronger link than simply that the listed artists share a spot on my list of unrelated favorites.
Jokes aside, having something show up on the page would probably work better.
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.]
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.
These services are clever, but I usually know all of their suggestions. :-(
And yes, I'm trying to do better.
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.
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!
Sometimes it works much better than stuff you get with bootstrap/UIkit/Foundation/whatever.
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
I searched "Don Caballero" on ifyoudig [1] and got a few good results, but also some very different artists, like Mogwai and Aphex Twin.
[1] http://ifyoudig.net/don-caballero
I think Mogwai and Aphex Twin are pretty good recommendations for someone who likes the math rock/post rock that Don Cab play.
Need any help on the front-end side of things?
ifyoudig.net seems to have a very small popularity bias compared to e.g. Spotify though, so that's nice.
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.