#genre: Part 2. Genre Tags

In the last installment of this #genre series, I laid out what I see as two competing narratives in popular music discourses: genre-is-dead and genres-are-(over)abundant. In this post, I’ll share one part of the results from my exploration of how these two ideas might coexist in some Spotify metadata. That is, I’ll show what kinds of genre labels Spotify gives to artists.

First, why focus on Spotify? It’s extremely popular, so it has a huge impact on how many people experience the ramifications of categorization. Second, they have an API that makes some of their metrics relatively easily accessible to the public.

Like any streaming service, Spotify has a bunch of ways they categorize their content and make recommendations. I’ll not get into their use of collaborative filtering, word2vec, or web scraping in this post, but you can check out this video if you want an idea of how some of this works. As Spotify gobbles up recommendation services, social media apps, database managers, and music AI companies, their black-box of categorization (and thus genre) becomes ever more opaque, its proprietary algorithms/data shrouded by their very heterogeneity.

But, different people working at Spotify have laid out part of the ideology behind their classifications. David Brackett explains:

In industry-based practice, Glenn McDonald’s work with EchoNest and Spotify illustrates some of the difficulties already discussed in connection to MIR [music information retrieval] work in general in its tension between trait-based reification and discourse-based folk taxonomies that guide quotidian use of genre labels. McDonald has explained that EchoNest’s response has been to rely on connections between artists rather than individual songs or albums as a way of organizing the similarity relations on which the company’s taxonomies are based (2016, 325).

So, in addition to all the recommendation processes that they base on your listening habits, Spotify has a layer of categorization that centers on artists. This strategy makes some sense; if they taxonomize artists, perhaps it will be easier—and more efficient for Spotify’s algorithms—to categorize the musical objects (tracks and albums) associated with those artists. This also means that Spotify acts like that stereotypical musicophile friend: oh you like Band X? You’d love (more obscure) Artist Y!

To categorize artists then, EchoNest/Spotify essentially created two techniques: genre labels and related artists, the latter of which will be the subject of my next installment. Each artist in Spotify is given a set of stylistic tags that seem to be largely curated through human intervention, while related artists are mostly determined by user-interaction and social media parsing. Unlike related artists, genre tags are hidden from listeners and users, but they permeate the platform. As of this writing (10/9/17), Spotify has 1520 unique genre tags, as glimpsed in this list. That’s a lot more than the bins at a record store.

(Spotify also stores the number of followers and popularity for each artist, and has a bunch of parameters that define the audio of individual tracks. A few things about this irk me as a music theorist, like key and mode being both separate and non-inclusive, energy being a quantifiable characteristic that spans genres, etc. Maybe I’ll look into those more critically another time.)

Now, any time people seek to categorize or quantify other people, there’ll be problems.  And the music industry in particular has a long track record of well-known shady racist and sexist classifications. Karl Hagstrom Miller describes how Plessy v. Ferguson, the American Folklore Society’s belief that music expressed stages of social and biological Darwinisms, and rapidly changing technologies of music distribution all combined to segregate musical objects along with musicians in the late 1800s. By the 1920s, “there existed a firm correlation between racialized music and racialized bodies” (4) which directly affected commercial viability of records and sheet music. In the late twentieth century, similar forces of hip-hop canonization bracketed out women as the the mainstreaming of rap was seen as an “emasculation.” Writing in 1990, Tricia Rose summarized the bleak situation: “for media critics generally, it is far easier to re-gender women rappers than to revise their own gender-coded analysis of rap music” (111).

Addressing current iterations of this legacy, Briana Younger’s recent piece presents compelling first-hand accounts of the direct effects that genre labels have on black artists today, and you should stop now and read it. Drawing on Jennifer Lynn Stoever’s new book, Younger describes black musicians’ navigation of stylistic expectations. It’s not just that people will label someone like KAMI as alt-rap. It’s that these labels bear directly on the opportunities available to and the mobility of these artists.

As an aside, two points in Younger’s title reminded me of Derrida’s influential essay, “The Law of Genre,” though these points run through lots of scholarship and popular writings on genre. Among other things, Derrida suggests two ways of understanding genre: as belonging (being “boxed in”) and as participating (“we fit in so many things.”) These conceptions intertwine in Younger’s article, as the musicians she interviews all struggle against agencies that limit their potential acts of participation. Moses Sumney laments that “genre is still relevant to people who run companies, and they still have power when it comes to music. The ability of a company to define genre is an exercise in power, and power is inextricably tied to economics.” Genre isn’t so dead.

Mapping and understanding genre tags

My general question in this part of the blog series, then, is, what happens when Spotify applies genre tags to artists? How does this large company activate conceptions of genre, and what effects do these genres have for musicians and listeners? A few examples of the kinds of genre tags Spotify applies:

  • Moses Sumney: deep indie r&b, escape room, indie r&b
  • Rihanna: dance pop, pop, r&b, urban contemporary
  • Taylor Swift: dance pop, pop, post-teen pop
  • Kendrick Lamar: hip hop, rap, west coast rap
  • Ed Sheeran: pop
  • Beyoncé: dance pop, pop, pop rap, r&b
  • Sam Hunt: contemporary country
  • The Beatles: british invasion, classic rock, merseybeat, protopunk, psychedelic rock, rock
  • Esperanza Spalding: contemporary jazz, contemporary post-bop, cool jazz, indie r&b, jazz, jazz fusion, neo soul, soul, vocal jazz
  • Charlie Parker: adult standards, bebop, big band, contemporary post-bop, cool jazz, hard bop, jazz, jazz blues, jazz fusion, soul jazz, stride, swing, vocal jazz
  • Dirty Projectors: alternative dance, alternative rock, anti-folk, brooklyn indie, chamber pop, chamber psych, chillwave, dance-punk, dream pop, escape room, folk-pop, freak folk, indie folk, indie pop, indie r&b, indie rock, indietronica, lo-fi, neo-psychedelic, new rave, noise pop, noise rock, shimmer pop, singer-songwriter, stomp and holler
  • Dr. John: acoustic blues, blues, blues-rock, boogie-woogie, british blues, chicago blues, classic funk rock, classic rock, country blues, country rock, deep funk, delta blues, electric blues, folk, folk rock, funk, jam band, jazz, jazz blues, louisiana blues, mellow gold, memphis blues, memphis soul, modern blues, new orleans blues, piano blues, psychedelic rock, rock, rock-and-roll, roots rock, singer-songwriter, soul, soul blues, southern rock, southern soul, texas blues, traditional blues, traditional folk

I chose these somewhat randomly to show the variety in both type and number of labels that artists get, and we can see that genre acts less like a boundary-defining noun and more like an adjective, used in combination for description. But this little list begs the question: how typical are these sorts of cardinality-discrepancies?

Example 1 below shows how many genre tags Spotify grants to a sample of about 200 artists. (Don’t worry, I have larger datasets and sample sizes coming.) The x-axis shows the number of genre tags, and on the Y-axis, we have their popularity. The graph contains a collection of 200 artists from rock, rap, and pop.  You’ll notice that the most popular acts, like Rihanna, and the least popular acts have relatively few genre tags; those acts with the most genre tags tend to be moderately popular. This Gaussian distribution—N.B. I know my fit line is not Gaussian, I’m still learning Tableau—makes a bit of sense for some of Spotify’s goals; they want you to “discover” new artists, and they help themselves by providing lots of genre-tag info for the artists that are not so well-known in hopes that they’ll be able to accurately predict if you’d like the music in this liminal range.


Example 1: Popularity vs. Number of Genre Tags for some artists in Spotify

But, let’s see what happens when we distinguish between artists associated with different large-scale categories or metagenres (that I have identified, so yes, there are some biases in my selections):


Example 2: The same as Ex. 1, now with rock, rap, and pop distinguished by color and shape.

The social, cultural, and economic capital represented by these genre tags gets distributed to artists rather unevenly. Clearly, rock is privileged in this instance. Now, of course some artists participate in many genres throughout their careers or within single albums, while some stay close to a generic center of gravity. The number of tags doesn’t measure quality, but, at levels above anecdotal evidence, genre tags essentially constitute measures of aesthetic value. If Spotify’s going to come up with 1500 genre labels, it’s worth understanding what kinds of musicians Spotify thinks deeply enough about to generate its stylistic descriptors. So, let’s see how this plays out at a much larger scale.



Example 3: 1005 Spotify genres and their interrelationships

Example 3 represents genre tags and their connections between about 12000 artists. Each node in the network is a genre. If genre A and genre B are both used on a single artist, then they get connected. The more artists tagged with those two genres, the thicker that line will be. I’ve emphasized the connections in this instance rather than the genres themselves. By letting Gephi automatically assign communities based on how the nodes interact, we get about 10 large genre-constellations, which I’ve labeled with via Paint (RIP).

(Yes that blue grouping with classical and techno is kind of odd, but it shows that Spotify thinks there are lots of connections between artists like Aphex Twin and Philip Glass. The tag, “intelligent dance music” shows their cards; you can probably guess the commonalities its musicians might share with classical music.)

Now, there’s a lot to unpack here that I’ll leave for future posts. But, just a couple of quick notes to explain how these graphs relate to Younger’s piece and to my conception of #genre. First, in Example 4 below, each node still represents a genre, but now the size of the node represents how many times that genre was used in my corpus. As you’ll notice, the dark green hip hop and pop cluster nodes are at least as large as the yellow and lavender genres. This means there are lots of musicians that participate in hip hop and pop in Spotify’s metadata.


Example 4: Node-size determined by frequency of genres, or how many artists receive the tags.


But what happens when we instead filter for how important these nodes are in the network, or how central they are when compared to other nodes. Are they linked to a lot of other genres? In other words, how mobile are artists that get tagged with these common genres? As you’ll see below, most of our green circles fade almost to imperceptibility, except for a small handful, in order of size: dance pop, pop, pop rock, funk rock, classic funk rock, and soul. Lots of pink and yellow nodes now surpass the hip hop nodes. Things like neo-psychedelic, alternative rock, electronic, indie r&b, rock, and singer-songwriter are all “more important” to the network than pop rap, urban contemporary, r&b, and hip hop. Artists tagged with these latter labels get boxed into their Local Group of Spotify’s genre universe; if you’re hip hop, you’re probably not much else to Spotify.


Example 4: Node-size determined by importance of genre tags in the network (by Pagerank).

For indie-pop artists in the lavender range, the homogenization of their genres’ network-centrality reinforces notions of musical omnivorousness, or the flattening of genre’s importance. If I’m tagged with indie pop, I’ll also likely participate in at least a dozen or two other genres, so each one doesn’t do a ton of adjectival work by itself. These are the kinds of artists that can fairly claim that genre doesn’t matter for them since they get an overabundance of stylistic labels that lessens the meaning of any single one of them.  But for those artists working in the green hip hop and r&b realm, these labels mean more and restrict their mobility.

Robin James’s study on the relationship between concepts of “post-identity” and “post-genre” provides similar conclusions. As James explains, “claims to genre transcendence are credible when they are made by artists who … appear free of any particular social identity. In order to sound post-genre, one has to seem post-identity. Beyoncé [or Rihanna] works all sorts of genres … but when the idea of her genre-transcendence is floated …  most people still interpret her as operating somewhere within R&B. … Only artists who inhabit the ‘non-black’ side of the post-identity color line … are legible as post-genre practitioners” (2017, 31). My post hasn’t really provided anything terribly novel in its results, but it shows how James’s convincing arguments play out in Spotify’s metadata, where genre mobility and cultural-capital investment intertwine.

I’ll post some more on my genre mappings at some other point, but my next post on #genre will tackle how artists relate and cluster together in Spotify and what the size and stylistic diffusion of these groupings might mean for how genre works. If you want to chat about this in person, I’m giving a talk on this stuff at SMT this year. So come on by!

P.S. Message me on Twitter if you’re interested in the datasets I used for this. I’m still working with them for some publications, but I’d be willing to share before I get those out. Some of my Python scripts are up on Github if you wanna mess around and get some of Spotify’s metadata yourself. Warning: it’s all clunky because I’m still less than a year into learning Python.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s