Monday, April 18, 2011

A 2D Map of the Music World Using Data from Last.fm


How does blues relate to rock, jazz, and soul?  What unites and separates different styles of music?  How greatly has the style of a band like the Beatles evolved over time?  These sound like questions for music experts, but as a fascinating new map shows, you may actually want to ask a group of five geographers and computing experts from Indiana and San Diego State Universities.

Creating the Map 

Music may not be a significant area of research for geographers, but among the many presentations at the annual AAG meeting in Seattle, there were at least a few dealing with music, two of which were of particular interest to me as a blues fan.  Sadly, I missed the session featuring the poster Poetics of Place from the Birthplace of the Blues: Blues Culture in the Mississippi Delta by John Byron Strait of Sam Houston State University, so I'll have to look out for him at next year's national AAG meeting.  However, on Saturday I was fortunate enough to catch a fascinating presentation of a map entitled A Semantic Landscape of the Last.fm Music Folksonomy, created by Joseph Biberstine, Russell J. Duhon, Katy Borner, and Elisha Allgood from Indiana University's Cyberinfrastructure for Network Science Center and Andre Skupin of San Diego State University.

As many of you know, Last.fm is a music website and social network that allows users to keep track of the music they are listening to and introduces them to other users with similar musical taste.  Users download the "Audioscrobbler" recommender system, which tracks which songs they listen to, both online and on their personal music devices.  Last.fm then suggests artists similar to the users' favorites and recommends users with similar musical tastes.

Like many sites, Last.fm allows users to tag songs with whatever they like, such as genre descriptions like "rock," facts about the song like "over twenty minutes," things related to the subjective experience like "seen live" or "my favorites," or even what the map authors describe as virtual graffiti like "altar of the metal gods."  This system of tags allows users to search for music by tags and creates a complex virtual network. 
Such networks of tags form what is known as a folksonomy, an intersubjective system of classification based entirely on how website users conceive of the tagged content (songs in the case of Last.fm).

The authors of this map have previous experience taking such complex, multi-dimensional data and representing it as a 2D map.  For this map they selected the 1,000 most used tags on Last.fm and used a form of a technique called a self-organizing map (SOM).  The basic concept is similar to the tag clouds formed by sites such as Wordle, in which each tag is sized relative to the number of times it appears, but it is much more advanced because it also attempts to place each tag near the tags to which it is most closely related.  The SOM forms an artificial neural network in which each tag is a "neuron" with connections to many other tags.  The SOM then analyzes the strength of the associations and pulls the data into a two-dimensional form.

What the Map Shows Geographically

The finished map shows a prioritized tag cloud in which each tag is assigned its little region of the map, bordering on the regions of tags it is most associated with.  Tags are ranked by how closely they are associated with a song.  For example, a song may be most commonly tagged "classic rock," but may also have the tags "blues rock," "70s," and "guitar."  Purple labels show the strongest associations and are separated by borders on the map, while secondary associations (and so on) are shown as sub-regions without marked borders and are labeled in different colors.  Some labels crop up in multiple places due to different associations and levels of association.  For example there are several primary regions labeled "female vocalists," as well as "female vocalists" sub-regions within other genre regions.

Although the space this map shows is purely intellectual, it is easy to think of it like a real space.  I like to think of it as a series of political and cultural entities.  Unlike continents and landforms, such entities are to an extent intellectual constructs, just as musical genres are, and are constantly mobile in the same way as well.  Each primary region is like a country, with each sub-region as a cultural or political enclave.  The genre- and style-related tags are more like cultural units, while the subjective tags, such as "seen live" or "my favorites" are more like political ones.

What the Map Tells Us about the Blues

One thing Biberstine and Skupin said surprised many people was the location of the blues.  As one would think, since most people don't know much about the blues, they expected to find its primary region over in the southeast corner of the map near jazz and soul.  This makes perfect sense because the blues grew up right next to these styles.  But many were surprised to find the blues is way over in the southwest corner, surrounded by various styles of rock.

This shouldn't surprise Blues Power fans!  As Muddy Waters said, the blues had a baby, and they named it rock and roll.  So of course the blues is surrounded by the styles of rock which most directly grew out of it, including the obvious one, blues rock, as well as the tags "classic rock," "progressive rock," "guitar," "guitar virtuoso," "garage," and "psychedelic."  However, because the map allows tags to appear in multiple places, the blues is also a major secondary level region straddling the border between jazz and soul.  So the map allows the blues to show up in both locations one would expect, even though the locations are widely separated by distant descendants such as metal and hip-hop.

The suprise at the location of the blues is definite confirmation of my opinion that far too few people know about the blues.  In addition, the primary region of the timeless and wonderful blues is dinky compared to those of most of the other genres, and even some rather ephemeral subgenres, such as shoegaze, and ridiculous pseudo-musical genres, such as dark ambient.

What the Map Tells Us about Last.fm

In his presentation, Biberstine said that many of the people who view this map think it is a fair representation of the musical world.  However, as admitted, since the sample "only" consists of 50,000 Last.fm users, it is far from definitive.  I think the map does an excellent job of showing the relationships between styles, but I'm not sure that it accurately represents the size of each style in the greater world as opposed to the Last.fm community.

For example, I think of hip-hop and rap as being a huge portion of today's music scene.  However, hip-hop (also spelled hip hop and hiphop in the Last.fm folksonomy), rap, and rnb (I'm assuming it's mostly the new stuff, that is all r and no b) take up about the same sized portion of the map as ambient and dark ambient.  What's up with that?!

So from this map you get more of an idea of the strength of various communities within Last.fm than across the entire world.  I would say that Last.fm has a ridiculously large community of fans of electronic, ambient, and industrial music, which takes up the entire northeast quarter of the map.  I don't like to be offensive, but seriously, why do people listen to that stuff when they have options such as...anything else?  There is also a surprisingly large community of metal fans in the south-central part of the map, which is fine by me because at least these guys can appreciate virtuosic guitar and still have a tenuous connection to the blues.  The pop community isn't as big as one might expect, but that's probably because pop is such a star-based industry that there are fewer songs and artists represented even though they are getting tons of listens.  It would be interesting to see another map in which number of listens was factored into the size of an area rather than just number of songs.  Other than that, most of the genre communities are of a seemingly proportionate size, with one notable exception.

It seems that hardly anyone on Last.fm likes country music.  As a result, it gets stuck in a very strange place.  At first, I couldn't even find it.  Traditional country is closest to folk and also has some connection to the blues, but the popular country nowadays is much closer to pop or rock.  However, in this map country can be found sandwiched between death metal and rap!  About the only explanation I can think of is that rap and death metal didn't quite go together, and country didn't go with anything else, so it just got stuck there.  Although maybe I'll be proven wrong by the forthcoming collaborative album from Taylor Swift, Snoop Dogg, and Six Feet Under.

How This Map Can Be Used

In addition to a conversation starter and great blog post material, this map is an excellent visualization of how a recommender system works.  Biberstine said that eventually the authors would like to work with Last.fm to incorporate this map into the site.  This would be really cool because rather than just following a bunch of links, users would have a single visual representation of where they were on the map and what styles of music they might want to explore next.  In fact, this approach may be useful in the future for other websites that combine tagging and recommender systems.

An audience member asked about the possibility of overlaying demographic data to see what types of people fell where on the map.  I pointed out that in the age of highly targeted internet marketing, this could be a powerful tool for identifying potential customer groups.  Skupin observed that there already are programs that allow this type of analysis, but creating a tangible map makes this information accessible to a wider range of people who might not have as much technological know-how.

As well as the two dimensions which can be seen right away, this map can also be used to show how artists and styles have evolved over time.  One could see which tags were associated with works from different periods to see if they would appear in different places on the map.  Skupin gave the example of looking at the tags associated with the songs on various Beatles albums, which would likely start out in "classic rock" and move into "psychedelic."

I'm sure there are many other cool applications for this type of folksonomy mapping that none of us even touched on today.  I'm sure that eventually these maps will pop up online on sites like Last.fm, Amazon, Google, Wikipedia, and others, and that we will have a lot of fun playing around with them.  Maybe, a number of years from now, there will even be a free online application like Wordle for designing simple versions of these maps.  However, according to Biberstine, the computations for this map took twelve days with parallel computing, or five months with linear, which is a bit longer than most of us would like to sit at our computers at any one time.  For now, the best bet is to print out a poster-sized version of the pdf, hang it on your wall, and keep rockin' out to your favorite tunes.

No comments:

Post a Comment