This process of fragmentation represents a Deleuzian distopia made real. The desire to translate the ephemeral ideas of musicality, and why certain music attributes appeal to certain people, The goal of the suggestion algorithm is “universal modulation,” a method of translating phenomena into a readable data, thus achieving an approximation of the original idea that can be quantized, marketed, and capitalized upon. Gilles Deleuze introduces the idea of humans as “dividuals,” entities comprised of multiple parcels of data that can be bought, sold, and traded in the new marketplace. In previous centuries, the individual was “one of many,” a member of a larger mass. Today’s dividual is known by his password, a code that can grant or restrict access to information¹. This new conceptualization recasts the each of us as a user, an aggregate of recorded preferences, histories, and proclivities. Every interaction with the market is thus mediated by the password: the human existence is constantly reconfigured in accordance to the user’s associated dataset.

However, this fragmentation process does not have to be viewed so pessimistically. Theorist Yochai Benkler posits that, as it applies to market architecture, fragmentation (or decentralization) can be healthy. Benkler described this sea change in a talk on “open-source economics” given at the 2005 TED conference. The industry can survive this dramatic shift by employing adaptive strategies that harness this threat.

Click here to view the embedded video.

A recommender system employing methods of collaborative filtering is an example of such an interface: each user provides invaluable data, and thus effectively markets the products to other users. The system provides the framework to sponsor a conversation about music between music enthusiasts. Users, as they share media, are willing to put in collaborative effort to make the web a better, more efficient place. For example, as the user base of Last.fm grows in number, so grows the number of linkages between users, and thus the opportunities to encounter new products increase.

• The Numerics of Style: Individuation and Suggestion Algorithms (1) |
• The Numerics of Style. Part 1: “Panic Algorithm” (4) |
• The Numerics of Style. Part II: “The Problem is YOU!” Function U, Explicit and Implicit Preferences, and Harlo’s IDM Ghetto (1) |

Recommendation systems vary one from another, but they all revolve around the following scenario: Imagine a digital vendor that has a finite set of items to sell (set S) and a potential consumer base (set C). Upon visiting the vendor’s website, a user should be directed to a subset (R) of items of high “usefulness” to him. The vendor’s server must employ an optimization function (function U) that, according to how it’s authored, is capable of generating an ordered subset of items, ranked by their usefulness, “on the fly.¹”

u: C x S → R

Crafting the fastest, most useful, and most personalized function U has been the Holy Grail in suggestion algorithm science, and in a relatively short span of time, the pursuit of function U has led to some hugely ingenious discoveries about the direct relation between computer learning, human psychology, and marketing in the entertainment industry.

At the outset, the recommendation system must gather an amount of sample data regarding a user’s preferences; this initial dataset will function as a focal point for function U to run successfully. This data is gathered in two ways: (1) explicitly, where a user is asked by the system to rank a number of items along a scale, and these answers are saved and “remembered” by the system every time the user comes back to the site, and (2) implicitly, where the application doesn’t ask the user, but closely monitors user interaction with the site to estimate the user’s preference². Although it is common for recommender systems to take both types of data into account, certain services privilege explicit user ratings over implicit user ratings.

Take for example, the interactive internet radio service, Pandora. When a user first visits the site (after inputting basic information such as username, email address, and zip code for account set-up), she is prompted to type in an artist they’d like to hear more of.

The first screen on Pandora Radio

The initial song chosen does not exactly typify the artist’s musical style, but rather certain musical qualities associated with that artist. In the example used here, UK grime (a rap subgenre) artist Dizzee Rascal starts with the track “Temptation” from his third album, “Maths + English” (2007). Clicking “Why was this song selected?” from the song’s menu button reveals that the song was played for the user based on her assumed preference for a Dizzee Rascal song that had “rock influences, basic rap roots, danceable beats, syncopated beats and swingin’ beats.” Because Pandora does not have much data from the user regarding the precise musical attributes she loves in Dizzee’s songs, Pandora chose to start with a song that had the most in common with Dizzee’s biggest breakthrough hits: The second single of Dizzee Rascal’s career, “Fix Up, Look Sharp” from his first album “Boy in da Corner” (2003) became a top 20 single, and garnered enough acclaim for the artist to win the coveted Mercury Prize for best album³. Although “Boy in da Corner” is considered to be a seminal work in the UK grime subgenre, “Fix Up, Look Sharp” was the watered-down crossover hit that introduced grime to the pop music charts.

Suppose the user in this example doesn’t want to hear toned-down grime—she wants to hear UK grime in all its fury. The user must then offer Pandora more explicit data: each track also takes a rating, giving the choices “I like it” or “I don’t like it.” Selecting “I don’t like it” causes Pandora to launch a new song from the artist’s catalogue.

The next Dizzee Rascal song, “Bubbles,” exemplifies other attributes usually associated with the artist: “club rap influences, electronica influences, funk influences, basic rap roots and beats made for dancing.” Here, Pandora tries the problem from another angle by offering a track that emphasizes other aspects of Dizzee’s music that made him such a popular, and marketable, artist. “Bubbles,” also from “Maths + English,” is stylistically seeded with attributes and techniques typical of that moment in rap/pop history. Pitchfork magazine quasi-laments this song’s artful parody of Southern (US) club rap, and accuses Dizzee of a miscalculated attempt to go “Trans-Atlantic” as he’s no longer relevant on the UK grime scene⁴. The user clicks “I don’t like,” and Pandora moves on. After feeding the application some more explicit data, she finally comes to a track she really loves: “Full Effect” by Tinchy Stryder, a lesser-known UK grime artist, whose sound embodies what the user seeks in the genre, and what she wants in a radio. Pandora will remember that this user has expressed explicit preference for songs with “hardcore rap influence, gangsta rap influence, and electronica influence.” She clicks “I like it,” and the next song is another she loves: “Jezebel” by Dizzee Rascal—a song similar to the previous in that it contains hardcore rap and electronica influence. Subsequent offerings by the station are equally “good” or useful to the user, for the most part.

Last.fm, however, places more value on the implicit, or user-suggested, item rankings. In order to gather implicit data, the application must have a method of tracking user interaction with the site, quietly observing the user to predict the usefulness of any item in the set of available products. To this end, Last.fm comes with a plug-in called the AudioScrobbler. The AudioScrobbler logs the user’s musical habits as she plays music on media player applications, such as iTunes, or even as she is untethered to the computer and listening to music on her iPod. These music logs are periodically sent (or “scrobbled”) to Last.fm’s servers to create a set of data from which the application makes its recommendations. The scrobbling practice is how the Last.fm service obtains implicit user data.

The Last.fm AudioScrobbler works with iTunes.

Where Pandora’s recommendations are governed by explicit data proffered by the user, and that data’s appropriate link to the music database, Last.fm’s recommendations are based on the user’s proximity to other users of the service. Last.fm offers the “Neighborhood” as a starting point: an ordered list of users whose musical tastes are similar to the user based on commonly-listened-to artists. In this example, a user’s Neighborhood contains 60 other users who share enthusiasm for the same sampling of artists. Using the method of collaborative filtering, Last.fm can assume that these musical neighbors’ tastes are so similar to the user that a high proportion of the neighbor’s music will appeal to the user.

The user’s Last.fm Neighborhood

Two of the user’s “neighbors,” with whom she shares uncannily similar tastes

The user is introduced to a variety of songs through the Neighborhood Radio, and for the most part, these songs are stylistically incongruous: In this example, the same user who ♥s Dizzee Rascal, was offered an “indie, synthpop” song by the band Passion Pit, then a more retro-sounding, French-pop inspired song by The Outrunners, followed by a lesser-know selection by jazz great Herbie Hancock. Last.fm’s Neighborhood Radio is concerned not with sending the user similar artists, or songs with similar musical attributes, but offering a subset of items most fitting to the type of listener that user represents. The user as a type of listener is best determined by the social group she belongs to. This principle is not unique to recommender systems; collaborative filtering is fundamental to search algorithmics. It is based on the establishment of user stereotypes, and employs heuristic artifacts that add extra weight to each item’s ratings. Heuristic artifacts take into account the degree to which a given set of users “like” an item: they exploit the standard deviation of an item’s ratings around a particular factor, most likely the genre that item ascribes to⁵.

My concluding installment will revisit the suggestion algorithm’s stake in the marketing of long-tail products, as well as evaluate the socio-political implications of these two approaches. Any comments are welcome!

• The Numerics of Style. Part III: (in)Dividuals (5) |
• The Numerics of Style: Individuation and Suggestion Algorithms (1) |
• The Numerics of Style. Part 1: “Panic Algorithm” (4) |

Question:
The first part of my investigation into suggestion algorithms answers the following question: How did the suggestion algorithm emerge as the culture industry’s main weapon in the fight to retain its legitimacy?

As our current era is written into tomorrow’s history books, scholars have only begun to contemplate the grave challenges digital technology has posed to the capitalist mode of production in the culture industry. The march of technology has effected drastic change in the ways art objects are produced, reproduced, and distributed. In doing so, technology has altered not only the underlying infrastructure in place to mediate the means by which these objects find their audience, but the very ways in which the audience apprehends art as an object.

So goes the Marxist axiom: an art object has inherent value that determines its price when brought to a market for consumption. In the traditional capitalist market, this price is established by calculating several factors: the “genius” of the artist; the mehrwert, or surplus value in the form of alienation, the process extracts from the artist; the actual costs involved in production of the work; and the cultural premium the potential customer will place on the exclusive ownership of the work ¹. The market does not have room for every product, so relatively few products can be relegated to the status of art object. This unhappy calculus has always presupposed physical scarcity: there is only so much space on a gallery wall, the FM spectrum only has so many channels, we only have so many hours a day to experience this object ². The culture industry has risen up over the past few centuries to mitigate this crisis of marketability and performs this very calculation to their profit. Modern digital technologies of production and reproduction force the predominating capitalist model to cross the Rubicon: the boundaries of physicality are eliminated and the formula must be re-evaluated in order to preserve the model. The culture industry must redefine their definition of the product in order to sell it.

Media theorist Kevin Kelly sounds the alarm in his essay, “Better than Free,” ³ advising industry heads to view the product in terms of its enduring qualities—immediacy, authenticity, and most importantly, its “findability” in the market. All music products are now technically identical: they are only ones and zeros floating in electric ether. The CD player, the mp3 player, or the internet radio widget cannot tell the difference between Coldplay’s latest hit⁴ and a cat bashing on a keyboard. Not only that, these two products cost roughly the same to reproduce and distribute, and to the eyes of some, have comparable artistic value. New distribution models must be created to deliver the “right” products to their intended audiences with these particular interfaces in mind.

The industry pushes the panic button. Enter the algorithm.

At their best, search algorithms allow a consumer to find exactly what he wants without even knowing what “that” actually is. The personalization layer operates here; a successful algorithm must sort through an immense repository of media based on the preferences of the user, and it must be able to respond to the preferences of any user at any given time. The goal here is two-fold: the algorithm facilitates the ultimate purchase, and the ease of use and the successful consumer-to-product match promotes a feeling of goodwill between seller and customer. This seemingly altruistic function is not the only purpose: the suggestion algorithm is the key to marketing niche products that would not likely find customers in the physical marketplace. It is often said that the potential for great profit lies beyond an immediate threat: the search algorithm is definitely poised to gainfully exploit this crisis in the entertainment industry through long-tail marketing of niche products.

The long-tail model is a relatively new contribution to marketing theory, promoted in 2004 by Chris Anderson, editor-in-chief of Wired Magazine ⁵. Anderson’s theory offers an optimistic view of the current market in contrast to the pervading sense of terror due to internet piracy, the rise of self-produced products (prosumerism⁶), and declining profit margins on physical media. Anderson’s model explains how offering a vast variety of products in the digital (read: non-physical) marketplace will actually encourage consumers to purchase more products, and in larger quantities. The model presents two major boons to the industry. It is abundantly easy to find markets for old or unpopular products: products that had been written-off as financial flops can find their appropriate audience. Second, new products that would never come to market for lack of an established consumer base are also a potential revenue source. The industry must then “shake hands with the devil”—embrace the small label, even the unsigned artist who produces out of her basement. Both products can now be sold with essentially zero additional opportunity cost to the distributor, and Anderson cites many success stories of companies that have employed long-tail marketing strategies to increase consumption of these niche products. Long-tail marketing necessitates the adoption of the suggestion algorithm. Powerful and accurate suggestion algorithms are the only thing needed to interface between the customer and the virtual warehouse.

I will begin my next installment with a deeper look at the processes involved in the suggestion algorithm, from a technical as well as a psychological perspective (including case studies from Pandora, Last.fm, The Echo Nest, and iTunes). Suggestion algorithms pull the customer along a chain purchasing options that are “automagically” bound together by a discrete set of criteria calibrated by the algorithm’s authors. We are all familiar with the phenomenon by now: “If you liked x, you’ll love y and z.” The optimization models employed to link y and z to your original preference for x varies from service to service, but the underlying principle remains the same: products we “like” to consume are typically those that both conform to a somewhat rigid set of aesthetic values and come with the stamp-of-approval of our peer group. Suggestion algorithms are successful in both of these regards. They parse the co-valences of the products surrounding us, and in the very act of suggestion, they implicitly recognize you as belonging to a social group revolving around certain sensorial/visceral ideals.

Stay tuned for more

• The Numerics of Style. Part III: (in)Dividuals (5) |
• The Numerics of Style: Individuation and Suggestion Algorithms (1) |
• The Numerics of Style. Part II: “The Problem is YOU!” Function U, Explicit and Implicit Preferences, and Harlo’s IDM Ghetto (1) |

In this travelogue, I will first discuss each platform’s technical approaches to data acquisition and methods of data analysis. These three examples employ radically different models, and various network topologies can be found in their differences. Next, I will move on to the broader topic of marketability. Each of these platforms have two sides to them: they are in service to public entertainment, but they also offer themselves as a marketing tool for the music industry. I would like to examine how this double-sided approach both radicalizes market research in the entertainment industry, and borrows from more traditional strategies. Finally, I will read these platforms with the Deleuzian concept of (in)dividualism in so-called “societies of control.” This module will examine how these three platforms exemplify the practice of “universal modulation” within our society, and how our particular capitalist age strives to turn such ephemeral human concepts into discrete packets of data.

Suggested readings can be found on my delicious under the tag “numerics_of_style” (and they’ll also pop up on the class’ TDMCC tag feed).

If anyone has any suggestions for readings for either the economic, technical, or philosophical thread of my travelogue, I would love, love, love the suggestion. And any critique along the way would be greatly appreciated. This should be a pretty fun, yet challenging, topic to tackle!Similar Posts:

• The Numerics of Style. Part III: (in)Dividuals (5) |
• The Numerics of Style. Part 1: “Panic Algorithm” (4) |
• The Numerics of Style. Part II: “The Problem is YOU!” Function U, Explicit and Implicit Preferences, and Harlo’s IDM Ghetto (1) |