Table of content
2. Historical development of recommender systems
2.1 Collaborative recommender systems
2.2 Content-based recommender systems
2.3 Context-aware recommenders: Exemplified by Spotify
3. Impact on listening practices
3.1 Discovering and curating practices
3.2 Mood management
3.3. The performativity of selfhood
4. Critical review
4.1. The commodified listener
4.2. Algorithms – the perfect structural listener?
4.3. Physiognomy of listening
“Die Flut präziser Information und gestriegelten Amüsements witzigt und verdummt die Menschen zugleich.“ “Dafür sorgt schon die Beschränkung der Information auf das vom Monopol Gelieferte, auf Waren oder solche Menschen, deren Funktion im öffentlichen Betrieb sie zu Waren macht.“
Adorno determines the explosive growth of information load as key feature of modern societies. Due to that, he questions the socio-cultural impacts of biased mass media. During the last decades digital music distribution solutions appeared, that changed the frameworks by which listeners access and experience music. According to “Bundesverband Musikindustrie”, digital music revenues (generated by downloads and streaming services) accounted for 38% of the market share in Germany 2016. The advent of streaming services and online shops established new opportunities to access enormous databases of information and cultural goods.
With the ubiquitous availability and rapid travelling of information, networked media environments confront consumers with an abundance of information, which they can not handle by themselves. As consequence of that, digital media platforms (such as Spotify) make use of recommender systems, that suggest items based on anticipated user preferences. Thereby, algorithms assist users to navigate huge databases of items. Recommender systems turn out to be one of the most powerful tools to cope with information overload. On the other hand, digital music platforms (such as Spotify) afford user with new opportunities to add, share, comment or rank items and, thus, facilitate the formation of participatory music communities. However, very little is known about the socio-cultural consequences of computer mediated decision-making processes.
Previous studies (e.g. Bull 2007, O’Hara and Brown 2006) mainly focused on the interaction of listeners with specific devices (e.g. ipod). Other strands spotted online fan activities (e.g. Burnett 2009, Bennett 2012) or commercial aspects of digital music industries (e.g. Jim Rogers 2013, Kjus 2016).
The following investigation aims to study the impact of context aware music recommender systems on habitual listening patterns. The underlying theoretical framework is based on sociological concepts and does not pursue psychological approaches. The purpose of the study is to outline:
- how recommender systems (exemplified by Spotify’s service “The Echo Nest”) influence individual and social encounters with music.
- how altered listening patters affect traditional notions of culture and its societal relevance.
The essay is based on the assumption, that music listening is driven by constant and stable motivational factors proposed by Schäfer: “People listen to music to regulate arousal and mood, to achieve self-awareness, and as an expression of social relatedness”. The investigation uses qualitative methods in order to answer the research questions. Findings concerning historical developments are based on literature research. Logical deduction and generalization are used to describe implications on listening patterns. Generalizations in chapter two relate to concepts provided by social action theories. All conclusions are evaluated in the light of critical theory.
Chapter two describes the historical development of music recommender systems. Key-features of context-aware recommender services will be exemplified by Spotify’s “music intelligence service”, called “The Echo Nest”. The following chapter questions the impact of music recommender systems on users’ listening habits. The chapter reviews and discusses findings of qualitative studies. The analysis of behavioral patterns does not intend to spotlight the formation and effects of music preferences. Chapter four takes a critical stance towards the socio-cultural and normative consequences of altered listening behaviors described in chapter two. The investigation is framed by Adorno’s theory of cultural industry. Even so, no references will be made to Adorno’s typology of listeners.
The ensuing considerations define “recommender systems” in line with Tarleton notion of “recommendation algorithms”. Later remarks dispense with further technological explanations. Gillespie characterizes algorithmic recommender systems based on three main criteria:
1. Patterns of inclusion:
the choices behind what makes it into an index in the first place, what is excluded, and how data is made algorithm ready.
2. Cycles of anticipation:
the implications of algorithm providers’ attempts to thoroughly know and predict their users, and how the conclusions they draw can matter.
3. The evaluation of relevance:
the criteria by which algorithms determine what is relevant, how those criteria are obscured from us, and how they enact political choices about appropriate and legitimate knowledge.
2. Historical development of recommender systems
The following chapter gives an outline of the historical development of music recommender systems. The illustrated systems assume that individual consumer choices follow stable and quantifiable patterns. In line with that, recommenders constitute clusters of potentially relevant choice-options. The relevance of choice-options can be determined through correlating habitual listening patterns with specific musical structures, demographic features of listeners (e.g. gender, age, ethnicity) or context information (e.g. weather, mood). Analytic categories proposed by Gillespie (“patterns of inclusion”, “evaluation of relevance”, “cycles of anticipation” ) will be used to exemplify how algorithms has altered music recommendation.
2.1 Collaborative recommender systems
“In everyday life, we rely on recommendations from other people”. Likewise, collaborative recommendation assumes that listeners who rate music similarly or show related listening behavior will react on other items similarly. Given a set of ratings, collaborative recommender systems aim to predict how listeners respond to an item that has not been rated yet. But how can listeners’ ratings be registered and matched with others?
As reported by Mark Levy, the first attempts to analyze mass media audiences (based on technology driven research design) can be traced back to the “Radio Research Project”, initiated by CBS in 1937. Frank Stanton, head of audience research for CBS, and the Austrian sociologist Paul Lazarsfeld invented the “Stanton-Lazersfeld Program Analyser”, a mechanical device being used to measure listeners approval rates. In experimental settings listeners were asked to press different buttons in order to indicate whether they liked bits of a prerecorded broadcast or not. The “Program Analyser” recorded and graphed listeners’ reactions against a time-line. Demographic data could be matched with user preferences by correlating listeners’ ”likes” with concrete musical structures. In 1938, Lazarsfeld convinced Adorno to participate in the project and promoted him as chief of the music division. After only two years Adorno quit. According to him, culture can’t be measured with reference to approval rates: “I reflected that culture was simply the condition that precluded a mentality that tried to measure it”. Nevertheless, ensuing “Program Analysers” were used by radio and television networks as well as advertising agencies until the 1980s. Refined approaches (like PEAC, the Voxbox or the Peoplemeter) introduced multi-variable response buttons as well as automated protocol systems (used to track demographic data: gender, age etc.).
Contrary to Lazersfelds and Stantons highly selective, sample-based laboratory approach, Nielsen developed a device called “Audimeter” that could be utilized to measure the listening behavior of mass audiences. The “Audimeter” monitored where and when the dial of a radio was switched. Every alteration of the dial button could be transcribed on a moving spool of paper. By matching those transcriptions with survey or diary data, specific listener profiles were constructed and sold by the “Nielsen Company” beginning in 1936.
One year later the Wurlitzer Company introduced the “playmeter”, which was utilized to measures the number of times a record was played by a jukebox:
“While weekly charts of record sales and radio playlists were influential, operators extolled the jukebox as a far more sensitive an accurate gauge of Americans’ musical taste. A customer who purchased a record for his home phonograph might play it until the grooves wore out, or play it only once; radio listeners frequently had to endure many lackluster songs and advertisement before hearing a favorite tune. The jukebox, by contrast, seemed to afford the consumer an opportunity to select exactly the song he or she wanted, and to play it once, twice or twenty times in a row…”.
Furthermore, jukebox owners had been able to quantify the aggregate tastes of local areas. Based on those information, regional radio stations began to develop independent top 40 playlists, addressing the preferences of local publics. The jukebox turned into the “most infallible device known to the music world for registering the public’s tune preference.”
With the emergence of digital collaborative filtering systems, technology based market research and recommendation shift from mechanical devices towards software solutions. Upcoming filtering systems maintain the former logic of “identifying people with similar interests and recommend items that have interested these like-minded people”. Therefor algorithms collect user rankings and provide further predictions, based on different filtering methods. In 1994, Shardan presented Ringo - one of the first internet services that implemented Collaborative Music Recommender System (CMRS). Ringo collected user judgments via email and provided algorithmic generated recommendation of albums and artists. From now on recommendations are no longer restricted by unlimited access to recordings (e.g. collection of record store, top 40 playlist of jukeboxes) and break with the narrow “patterns of inclusion” provided by analogue approaches. The online radio station Last.fm (,founded in the United Kingdom 2002,) relies on preference profiles built by listeners’ “likes” or “skipping- behaviors”. By that, algorithms allow listeners to shape personalized recommendations. So, music recommendation shifts from static prediction towards “cycles of anticipation”. Other CMRS (used by itunes or amazon) take into account which items had previously been consumed by the customer or related profiles. Thus, algorithms refine the “evaluation of relevance” by matching a multitude of ratings and listening biographies. As a result of this, clusters of potentially relevant choice-options can be diversified and adopted to listeners with specific demographical features (e.g. gender, age, ethnicity).
However, CMRS’s underlying rationale shows close resemblance to Bourdieu’s concept of “cultural capital”, claiming that listener decision-making is predefined by external and stable socio-economic factors: “… nothing more clearly affirms one's 'class', nothing more infallibly classifies, than tastes in music.” The following concepts challenge this apprehension, considering listening patterns (-not only preferences!) as being linked to specific musical structures.
2.2 Content-based recommender systems
Content-based filtering expect that specific “genes” or musical structures are preferred by individual listeners. It assumes that musical items can be analyzed and registered by a set of inherent features. However, how can music be structured according to predefined categories and matched with listening patterns?
In 1968 the “Institute of Contemporary Art” in London curated an exhibition “Cybernetic Serendipity” focusing the interaction between computers, art and audiences. Alongside of Xenakis and John Cage, Zinovieff exhibited a computer that played variations on tunes that you whistled to it:
“The computer analyzed the whistle and would often guess what the person was going to whistle next. I took several of the most popular tunes that people would whistle and the computer would make tunes out of the whistles”.
Consequently, Zinovieffs computer could be considered as technology based attempt to predict musical items with reference to consumer preferences. In spite of this, the computer was not able to analyze music in reference to predefined characteristics. A more systematically approach was provided by David Cope - former professor of music at the University of California. Cope developed several softbots (“Emmy” 1987, “Annie” 2012) aiming to compute compositions in the manner of classical composers. Despite “Emmy” and “Annie” had never been able to generate authentic musical content, new avenues could be explored to describe and register musical items based on the grammar of algorithms. As a consequence of that, huge amounts of songs could be organized in databases.
Whereas music analysis companies such as Poliphonic HMI uses algorithms to analyze and archive musical items, Pandora Internet Radio (founded as “Savage Beast Technologies”) employed hundreds of musicians to listen to songs and categorize them by more than four hundred different attributes. In contrast to record shops (just as amazon or itunes), conventional radio stations or collaborative filtering systems, Pandora classifies music according to musical “genes”. In 2000, Will Glaser and Tim Westergren (the founder of Savage Beast Technologies and Pandora) launched the “Music Genome Project”. The “Music Genome Project” system is a large database of records. It is associated with a set of search and matching functions that operate on that database. The “matching algorithm” calculates the distance between a source song and the other songs in the database and then sorts the results to yield an adjustable number of closest matches. That is to say, that after choosing a song, Pandora compares the song with the “genes” of every song in the database in order to recommend contents with similar attributes:
“Each gene corresponds to a characteristic of the music, for example, gender of lead vocalist, level of distortion on the electric guitar, type of background vocals, etc. In a preferred embodiment, rock and pop songs have 150 genes. Other genres of music, such as world and classical, have 300—500 genes.”
Pandora learns more about the users’ music taste by analyzing similarities of selected items as well as user ratings. Along the lines of Gillespie, “cycles of anticipation” allow listeners to influence recommendations. Hence, the “relevance of content” is evaluated by “raising the weights of genes that are important to the individual and reducing the weights of those that are not.”
2.3 Context-aware recommenders: Exemplified by Spotify
Founded in 2008, Spotify developed to one of the world’s leading music streaming platforms, providing access to more than 30 million songs to around 140 million users. Since 2015, Spotify affords users with context-aware music recommendations, generated by its music intelligence service called “The Echo Nest”. The recommendation algorithms are built on tradition filtering technologies (content-based or collaborative) as well as context aware metadata concerning listeners music behavior and daily-live habits.
Context-based filtering, however, assumes that contexts (daily life activities, emotions, social relatedness) influence listening habits: ”Das Auswählen und Hören von Musik [varriiert] heute sowohl personenspezifisch als auch situativ, das heißt inter und intraindividuell [...]”. Algorithms anticipate which music could be preferred by individual listeners in relation to certain contexts. However, listeners intrinsic intention of choice is not that relevant – what matters is the correlation between quantified context data and listeners’ reactions to them.
2.3.1 Patterns of inclusion
Ongoing trends of digitalized music distribution transform the embodiment and, by that, users’ notion of materiality: “This way of distributing music can, likewise, be seen as turning it from an artifact (LPs, cassettes, CDs) into a database resource that is always accessible.” In such way, private collections of “artifacts” were replaced in favor of databases. In the light of that, it is worth discussing the impact of algorithms (such as “The Echo Nest”) on the way music is presented to the listeners.
At first glance, databases represent data with no regard to hierarchies. Algorithms used to structure databases, open up endless ways to organize content. By that, databases could be categorized according to the structure of one’s own, long-cherished CD collection. As compared to the “private CD shelf”, algorithm based structures turn out to be different in some respects. Manovich argues that accessing Spotify databases based on algorithmic recommendations leads to an emphasis of information and requires a constant process of searching and selecting at the expense of narratives and contexts: “today we have too much information and too few narratives that can tie it all together.” According to Manovich, physical features of an album (e.g. autographs of artists on cover) are closely related to narratives (e.g. the remembrance of youth). According to that, meanings are “trapped within the commodity” and can not be converted to digital formats or computed by algorithms.
Chapter three will describe how listeners make use of recommender systems in order to create new narratives, cope with immateriality and develop new concepts of ownership.
2.3.2 Cycles of anticipation
Digital providers are not just providing information to users but also users to their algorithms. However, most platform make it their business to know much more about the user than the query listeners just entered. In so doing, music providers
“must not just track their users, they must also build technical infrastructures and business models that link individual sites into a suite of services (like Google’s many tools and services) or an even broader ecosystem (as with Facebook’s “social graph” and its “like” buttons scattered across the web), and then create incentives for users to remain within it.”
In the case of Spotify, “cycles of anticipation” determine the interaction between listener and recommender system. Thus, Spotify’s “cycles” involve context aware data generated by third-party applications (e.g. facebook or twitter). For instance, feedback-loops could be generated by smartphone-sensors, measuring how listeners respond to items in specific contexts. The (Spotify-) app “running” recommends music that fits to the frequency of steps while running. Other apps (such as “climatune”) produce “cycles of anticipation” in relation to weather data.
2.3.3 evaluation of relevance
When Spotify user ask for recommendations, “algorithms must instantly and automatically identify which of the trillions of bits of information best meets the criteria at hand, and will best satisfy a specific user and his presumed aims.” Further investigation of Spotify’s evaluation processes ask for a critical analysis of the algorithm to question its underlying criteria. However, Spotify does not provide any additional information concerning its algorithm’s workings. As a result, Gillespie argues that criteria of evaluation “skew to the provider’s commercial or political benefit”.
 Max Horkheimer, Theodor W. Adorno, Dialektik der Aufklärung. Philosophische Fragmente, Frankfurt am Main
2000, p. 11.
 Ibid., p. 351.
 Cf. Thomas Schäfer et al., “The psychological functions of music listening”, in: frontiers in psychology (2013),4, article 511.
 Tarleton Gillespie, “The Relevance of Algorithms”, in: Gillespie, Boczkowski, Foot (eds.), Media Technologies:
Essays on Communication, Materiality and Society, Cambridge 2014, p. 169.
 Tarleton Gillespie, “The Relevance of Algorithms”, in: Gillespie/ Boczkowski/ Foot (eds.), Media Technologies: Essays on Communication, Materiality and Society, Cambridge 2014, p. 169.
 Paul Resnick, Hal Varian, “Recommender Systems”, in: Communication of the ACM 40 (1997), p. 56.
 Cf. Rashmi Sinha, Kirsten Swearingen, “The role of transparency in recommender systems”, in: Terveen/ Wixon (eds.), Proceeding CHI’02 Extended Abstracts on Human Factors in Computing Systems, New York 2002, p. 830.
 Cf. Jonathan Herlocker et al., “An algorithmic framework for performing collaborative ltering”, in: Proceedings of 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR
’99), Berkeley 1999, p. 232 f.
 Cf. Mark Levy, “The Lazarsfeld-Stanton Program Analyzer: An Historical Note”, in: Journal of Communication 32 (1982), nb. 4, p. 33.
 Stefan Müller-Doohm, Adorno: A Biography, Malden 2005, p.247.
 Cf. Barrie Gunter, Media Research Methods: Measuring Audiences, Reactions and Impact, New York 2000, p. 149.
 Ibid, p. 150.
 Cf. James Twitchell, Adcult USA: The Triumph of Advertising in American Culture, New York 1997, p.95.
 Cf. Norman Medoff, Barbara Kaye, Electronic Media: Then, Now, and Later, Oxford 2011, p.188.
 Chris Rasmussen, “ ’The People's Orchestra': Jukeboxes as the Measure of Popular Musical Taste in the 1930s and 1940s”, in: David Suisman, Susan Strasser (eds.), Sound in the Age of Mechanical Reproduction, Philadelphia 2009, p. 189.
 M.G. Hammergren, “Is it a Hit? Ask the Automatic Phonograph”, Billboard 31.Jan.1942, p. 82.
 Rashmi Sinha, Kirsten Swearingen, “The role of transparency in recommender systems”, in: Terveen/ Wixon (eds.), Proceeding CHI’02 Extended Abstracts on Human Factors in Computing Systems, New York 2002, p. 830.
 Cf. Jonathan Herlocker et al., “An algorithmic framework for performing collaborative ltering”, in: Proceedings of 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’99), Berkeley 1999, p. 231.
 Cf. Upendra Shardanand, Pattie Maes, “Social information fitering: Algorithms for automating\word of mouth", in: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, Denver 1995, p.211.
 Elena Razlogova, “The Past and Future of Music Listening: Between Freeform DJs and Recommendation Algorithms”, in: Hilmes/ Leviglio (eds.), Radio’s New Wave: Global Sound in the Digital Era, New York 2013, p.68.
 Cf. Ibid., p.69.
 Pierre Bourdieu, Distinction: A Social Critique of the Judgement of Taste, 8th edition, Cambridge 1996, p. 18.
 Cf. Prem Melville, Vikas Sindhwani, ”Recommender systems”, in: Sammut, Webb (eds.), Encyclopedia of Machine Learning, New York 2011, p.834.
 Cf. Rainer Usselmann, “The Dilemma of Media Art: Cybernetic Serendipity at the ICA London”, in : Leonardo 36 (2003), nb. 5, p. 390.
 Peter Zinovieff, “The Russian-English Renaissance Man’s Guide to Quadrophonic Sounds“, 20.08.2017,
retrieved from: http://redbullmusicacademy.com/lectures/dr-peter-zinovieff-the-original-tectonic- sounds?template=RBMA_Lecture%2Ftranscript.
 Cf. Christopher Steiner, Automate This: How Algorithms Took Over Our Markets, Our Jobs, and the World, London 2012, p.99 ff.
 Cf. Ibid., p. 83.
 Ibid., p. 83 f.
 Cf. Consumer item matching method and system. U.S. Patent No. 7.003.515, 14.08.2017, retrieved from: http://www.google.com/patents/US7003515?dq=7;003515, p.7.
 Ibid., p.8.
 Spotify, 15.08.2017, retrieved from: https://press.spotify.com/us/about/
 Peter Vorderer, Holger Schramm, “Musik nach Maß. Situative und personenspezifische Unterschiede bei der Selektion von Musik“, in: Deutsches Jahrbuch für Musikpsychologie 17 (2004), p.90.
 Patrik Åker, “Spotify as the soundtrack to your life: Encountering music in the customized archive”, in: Johansson (ed.), Streaming Music, Oxon 2017, p. 81.
 Lev Manovich, The language of new media, Cambridge 2001, p. 217.
 Nicky Gregson, Louise Crewe, Second-hand cultures, Oxford 2003, p.144.
 Cf. Tarleton Gillespie, “The Relevance of Algorithms”, in: Gillespie/ Boczkowski/ Foot (eds.), Media Technologies: Essays on Communication, Materiality and Society, Cambridge 2014, p.173.
 Ibid., p.175.
 Ibid., p.176.
- Quote paper
- Anonymous, 2017, Impact of context-aware recommender systems on habitual listening patterns, Munich, GRIN Verlag, https://www.grin.com/document/381036