The digital realm is not what it seems, and impacts on music curation in myriad and hidden ways, says Tully Barnett, one of the lead researchers in the Flinders University team Laboratory Adelaide: The Value of Culture.
Barnett and Julian Meyrick and Robert Phiddian are co-authors of What Matters? – Talking Value in Australian Culture (Monash University Publishing, $24.95). Below is an extract from the chapter “Digital Disruption/There’s an App for that!”
Disruption is a loaded word. But it is also an empty signifier, hollowed out by misuse and overuse, consumed and regurgitated by corporations hungry for the next slick management term. As we go to press, happenings in this field are fluid. It is too soon to tell how the #metoo movement and the 2018 Facebook/Cambridge Analytica data sharing scandal (and the algorithmic reality underpinning it) will affect our lives long-term.
Digital disruption, however, means something more particular. Wikipedia (a senior member of the Digital Disruptors’ Club) says that in the field of business the term refers to ‘an innovation that creates a new market and value network and eventually disrupts an existing market and value network, displacing established market leading firms, products and alliances’.
The phrase was coined by Clayton Christensen in his 1995 book The Innovator’s Dilemma. But the term has mutated in usage, as terms tend to do. In the NPR program All Things Considered, Kevin Roose points out: ‘these days [disruption]’s used to sort of mean cool … [and] anything that’s sort of vaguely new or interesting’. But the word ‘digital’ needs some investigation too. It is just as ubiquitous, if seemingly less controversial.
Let’s take a quick tour around the major digital disruptors:
Spotify, the music streaming service, grew out of a response to file-sharing practices, capitalising on the early failure of the less-than-legal Napster, established in 1999 as a peer-to-peer file-sharing platform. Where Napster had no connection with the artists whose work it distributed, Spotify paid royalties to its musicians, albeit insultingly low ones.
Music lovers rejoiced to find a convenient and responsible way to listen to old favourites and discover new ones. For many of its users, it is Spotify’s recommendation engine that makes the subscription fee attractive. Again, this engine employs algorithms that note what you are adding to your playlists, what you are listening to and, crucially, what you are skipping over, to shape a suite of 30 new songs, a customised mixtape ‘for your listening pleasure’, once a week.
There are many different versions of recommendation engines, employing different approaches to the ‘value-add’ role of curation or discovery. Think about Amazon’s prompt: ‘people who bought this also bought …’. Sometimes it’s useful, sometimes it’s hilariously dumb. It’s a crude system relying on the punt that similar-seeming customers will have similar interests. When Spotify’s metadata style guide was leaked in 2015, it revealed the usual technical advice: how to deal with different or non-standard spellings of a name; how to account for creative roles (including producer and lyricists, as well as performing artists); the problem of remastered releases; the categorical distinction between a single, an EP, an album, a compilation; and so on. But in doing so it also released a lot of less innocent information about their techniques for generating recommendation lists.
Pandora is Spotify’s best-known antecedent, though there is also Last.fm, and Apple Music is currently seeking a stronger market foothold. Pandora’s curation depends on tagging music by attributes. Its Music Genome Project ‘captures the essence of music’ by reducing music to 450 attributes, or ‘genes’, via an in-house team of musicologists. These musicologists listen to 20–40 seconds of a song then attach metadata, a list of relevant attributes, to classify it.
Why do women artists appear less frequently than men in the recommendation list?
Sub-genomes determine the fields to be populated (a folk music song will generate a different set of possibilities to swing or heavy metal). The attributes count some things that can be measured precisely: beats per minutes, use of particular harmonies or instruments, etc. Other traits are less objective, such as ‘musical influence’ or how dominant a rhythm is or the intensity of a track. There is training for this, calibration, peer review. But in the end it is what it appears to be: personal judgment.
This inevitable subjectivity raises inevitable questions about partiality. Why do women artists appear less frequently than men in the recommendation list? Are the reasons for this systemic or cultural? Is it because fewer women are played on the radio or get recording contracts, so fewer women appear in self-generated playlists, so fewer women appear in recommendation lists? One of Laboratory Adelaide’s research team tried to alter this, by adding only women artists for several weeks in a row, thus expressing a clear musical preference. But it didn’t have much effect on the recommendations arriving in the playlist each Monday morning.
Recommendation engines tend to be opaque for commercial reasons, which means that even though we know the result, we can’t discover what drives the choices. The engine in Spotify is a big data project that depends on and deploys our ‘taste profiles’, generated from our listening habits. These are correlated with the more than two billion playlists inside Spotify generated by its 140 million people, of which 70 million are paying users.
The Spotify team has made some of its technical information available through a Slide Share presentation, ‘From Idea to Execution: Spotify’s Discover Weekly’. According to this inside information, the big data of users’ playlists is then processed using collaborative filtering and natural language processing. Spotify treats a playlist as a document, and the songs in a playlist as words, and their team uses commonly available text mining tools to drill deep into the data.
Like Netflix, Spotify uses curation as a ‘value-add’. Users can both access and discover content through their services and the big data algorithms developed behind the scenes. Value in this context is conditional to the key terms applied to discover it, and the tail of metadata inevitably wags the dog of content. What needs to be much better understood, therefore, is the decisive impact of this hidden curation on our actual cultural experience.