Crossfade: Observing Transitions from Broadcasting to an
Algorithmic “Hot Clock”
Analysis of Spotify’s Content Curation Algorithms
Master Thesis University of Amsterdam Graduate School of Humanities Media Studies: New Media and Digital Culture
Author: Oskar Štrajn Student Number: 10849912 E-‐mail: oskar.strajn@gmail.com
Supervisor: Anne Helmond, MA Second Reader: David Nieborg, PhD Amsterdam, 26 June 2015
ABSTRACT
In a society in which digital technology is highly embedded in the everyday life, the processes of automated information forwarding represent a significant form of authority. By observing the current models of content forwarding that employ algorithms, we can identify that these systems are imitating structures already used by the traditional broadcast media. In this thesis, with the help of Spotify’s algorithms as a case study, I present a comparison of algorithms with the traditional FM radio “hot clock”; a schema indicating when particular music selections are to be aired. Furthermore, similar to when broadcast media profoundly influenced popular culture and everyday habits in the pre-‐internet, software on influences contemporary culture; the latter phenomenon shall be observed in this thesis.
ACKNOWLEDGEMENTS:
I would like express my deepest gratitude to everyone who encouraged, supported and assisted me to writing this thesis. I would also like to extend special thanks to my supervisor Anne Helmond for guiding and advising me on how to express my thoughts.
Mojca and Zmago Štrajn Blaž Blokar
Gabi Rolih Klemen Šali Aleks Jakulin, PhD
T
ABLE
O
F
C
ONTENTS
1.
INTRODUCTION 6
1.1.
CLOUD MUSIC SERVICES 7
1.2.
BRIEF HISTORY 10
1.3.
DIGITAL MUSIC UBIQUITY 11
1.4.
MUSIC STREAMING PLATFORMS ARE SOFTWARE 12
1.5.
SOFTWARE AND ALGORITHMS 13
1.6.
RESEARCH QUESTION 16
1.7.
OVERVIEW 16
2.
LOOKING INTO THE SOFTWARE 18
2.1.
THE ERA OF ALGORITHMS 18
2.2.
RECOMMENDATION ALGORITHMS OF STREAMING MUSIC SERVICES 21
2.3.
SPOTIFY’S RECOMMENDATION ALGORITHMS 22
2.3.1.
PERSONALIZATION ALGORITHM 23
2.3.2.
CONTEXT ALGORITHM 27
3.
METHODOLOGY 31
3.1.
INTERFACE ANALYSIS 31
3.2.
ALGORITHM ANALYSIS 33
3.3.
LIMITATIONS 38
4.
CONTEXT AND CURATION 39
4.1.
FRONT-‐END INTERFACE ANALYSIS 39
4.1.1.
FEATURED PLAYLISTS ANALYSIS 42
4.1.2.
ANALYSIS SPONSORED CONTENT ALGORITHMS 45
4.1.3.
INTRODUCING SPOTIFY RUNNING 47
4.2.
BACK-‐END API DATA ANALYSIS AND CONTEXT-‐CURATED RECOMMENDER 48
4.2.1.
ADVANCED CONTEXT-‐CURATED RECOMMENDER 53
5.
THE PAST IS THE FUTURE 55
5.1.
THE CURATORS 55
5.2.
MEDIA SOFTWARE AND TRADITIONAL BROADCAST STREAM 57
5.3.
MOVING BEYOND BROADCAST MEDIA 59
6.
CONCLUSION 61
7.
BIBLIOGRAPHY 64
8.
APPENDIXES 69
8.1.
APPENDIX 1 -‐ RESEARCH SOFTWARE CODE 69
8.2.
APPENDIX 2 -‐ RESEARCH SOFTWARE FULL DATASET 69
TABLE OF FIGURES
FIGURE 1 – ALGORITHMS DIAGRAM 14 FIGURE 2 -‐ SPOTIFY DISCOVER FUNCTIONALITY 24 FIGURE 3 – FEATURED PLAYLISTS SECTION 28 FIGURE 4 –PLAYLIST CREATORS 29 FIGURE 5 -‐ CREATION OF RESEARCH ACCOUNT 32 FIGURE 6 – SCRAPING ALGORITHM SCHEME BY SANDVIG AT. AL. 33 FIGURE 7 -‐ SPOTIFY CONTEXT TARGETING FOR BRANDS 34 FIGURE 8 -‐ SPOTIFY DEVELOPER APPLICATION 35 FIGURE 9 -‐ RESEARCH SERVER ROOT 35 FIGURE 10 -‐ VIRTUAL PRIVATE SERVER (VPS) 36 FIGURE 11 -‐ COLOR-‐CODED DATA IN EXCEL 37 FIGURE 12 -‐ MAINTENANCE ANNOUNCEMENT 38 FIGURE 13 -‐ SPOTIFY OPENING SCREEN 40 FIGURE 14 -‐ DISCOVER WITHOUT USER’S LISTENING HISTORY 41 FIGURE 15 -‐ BROWSING THROUGH SIMILAR ARTISTS 42 FIGURE 16 -‐ SECTIONS OF FEATURED PLAYLISTS 43 FIGURE 17 -‐ FEATURED PLAYLISTS EDITORS 44 FIGURE 18 -‐ SPONSORED PLAYLISTS SUBJECTED TO CONTEXT ALGORITHM 45 FIGURE 19 -‐ USERNAME CHANGE WHILE LOADING PLAYLIST 46 FIGURE 20 -‐ CHART GRAPH OF PLAYLIST GROUPS WITHIN THE RESEARCH WEEK 48 FIGURE 21 -‐ EXCEL COLOR FILTER 49 FIGURE 22 -‐ PARTY PLAYLISTS 50 FIGURE 23 -‐ GETTING READY AND COMMUTE PLAYLISTS 51 FIGURE 24 -‐ OVERVIEW OF THE DATA 52 FIGURE 25 -‐ CONTEXT DATA FROM 27TH OF APRIL 54 FIGURE 26 -‐ ROAD TRIP PLAYLISTS 58
1. I
NTRODUCTION
In a society in which digital technology is highly embedded in everyday life, the processes of automated information forwarding representing a significant form of authority. By observing the current models of content forwarding that employ algorithms, we can identify that these systems are imitating structures already used by the traditional broadcast media. In this thesis, with the help of Spotify’s algorithms as a case study, I present a comparison of algorithms with the traditional FM radio “hot clock”; a schema indicating when particular music selections are to be aired. Furthermore, similar to when broadcast media profoundly influenced popular culture and everyday habits in the pre-‐internet, software on influences contemporary culture; the latter phenomenon shall be observed in this thesis.
In recent years, the usage of cloud music services has rapidly expanded, promising a new era of listening. Like physical music media, digital ones have also experienced an evolution. Here the discussion is not only about the development of new file types (.mp1, .mp2, .mp3, .flac, etc.), but about vast music databases managed by a music streaming platform. When I am referring to online streaming media systems, I mean those that offer music as their service. In particular, I am looking into already established platforms, the aim of which is to replace direct online file exchanges. The databases that offer the direct access to content are called “cloud databases”. As all digital technology is moving towards cloud computing, music listening is also following this trend (“European Cloud Computing Strategy”). The first music streaming platforms that started to offer online streaming were introduced in the early 2000s, although they only recently have become widely used, as explained in his reviews of a music cloud service (Haupt 132). In this thesis, I also refer to the music cloud services with the terms “social music platforms” or “music-‐streaming platforms or services”. Cloud music services, such as Pandora, founded in 2000, offering the first personalized online radio, and the music social network Last.fm, founded in 2001, introduced themselves to the public as a modern supplement to radio stations, but without the possibility of being ubiquitous like the traditional radio. Nevertheless, they offered something completely new: a personalized experience of radio streaming (Haupt 132). Last.fm did not attain great commercial success, although it was very popular from 2006 and 2008. By the end of the latter year, it had made a loss of two million pounds (UK) loss, so they chose to close its radio service and focus only on music recommendations (Sweney). Despite Last.fm no being able to create a digital environment that would attract a great
number of users, a new cloud music service, Spotify, founded in 2006, has experienced significant success (Haupt 138).
Once that the first streaming music platform experienced success, other services also became interesting for users. As introduced in the IFPI Digital Music Report 2014, streaming and subscription platforms are currently overtaking the music market, mainly by global brands such as Deezer and Spotify (“Digital Music Report — IFPI — Representing the Recording Industry Worldwide”). There are a number of reasons streaming services have become a matter of interest for users. Streaming music platforms offer legal on-‐demand music, which only became possible after the Internet. Before, one would need to own all music releases on physical media to have the same experience. However, this was already possible in pre-‐streaming times, where music was available through illegal channels. Music streaming platforms here helped with legalization by paying the artists royalties and aiding in the organization of an extensive content database.
This thesis is concerned with the concepts of content “curation” being made by the platforms. Curation is a process of organizing and displaying information relevant to the user. Specifically, within the music streaming platforms, I address the curation of content that is displayed by software. In order for them to function, the music streaming services are using algorithms to present the information; due to the popularity of these services, algorithms are becoming influential in the creation of popular culture.
1.1. CLOUD MUSIC SERVICES
Cloud music services appeared as a result of the evolution of digital technology and have resulted in the creation of new ways of music consumption. The first services appeared in 2002 and introduced a new digital format that allowed listening music without transferring the files to a personal device. This can be seen as a next step in the development of music formats, from vinyl to track recordings, from cassettes to CDs and minidiscs, mp3 files, to online streaming. Streaming is still based on digital music formats, as the music streamed is saved in mp3, wav, or similar files. The main difference here is that the principle of music ownership has changed, as the files are not located on the user’s computer, but inside a digital cloud. The interest in the creation of music streaming services appeared due to the disruptions in the music industry created by online piracy. Streaming music platforms managed to create an environment that was capable of managing artists’
rights while simultaneously satisfying users’ expectations of free music on-‐demand. By free here I have in mind, add-‐supported music streams.
This means that any user is able to play any song at any time or place, as if the user would own the entire database. Another advantage that cloud services provided was the possibility of accessing the cloud database whenever an Internet connection was available. This advantage that was offered by music streaming platforms (the possibility of streaming music at any time or place) made this software as ubiquitous as traditional radio.
Cloud music services or streaming music platforms are a combination of two different fields within the creative industries (Ahvenniemi et al. 170). In the article Creative
industries and bit bang – how value is created in the digital age, the authors explain how the
music industry is connected with technology and why software also is a part of the new music industry (Ahvenniemi et al. 173). Although the fields are becoming so close to each other, I am not focusing on their relationship within this thesis. I decided to pay attention to the software industry, since the platforms are its creation. However, it must be acknowledged that the content that they offer is created by the music industry, and that as music streaming platforms are becoming one of the main distributers of digital music, they also have to be analysed from music industry’s perspective (“Digital Music Report — IFPI — Representing the Recording Industry Worldwide”). In order to do so, in some cases I acknowledge the music industry perspective to exemplify the motivation for software developments. Furthermore, when I am talking about the music industry, I am referring to the music rights holder and, in the most cases, I introduce their interests and views with the usage of statements presented by the International Federation of the Phonographic Industry (IFPI). IFPI presents itself as the voice of recording industry worldwide since it represents the interests of 1,300 record companies from across the globe (“About — IFPI — Representing the Recording Industry Worldwide”). As John B. Meisel and Timothy S. Sullivan noted, the music industry, which is institutionalized and in that sense traditional, has been forced to reorganize due to the Internet and digital developments (22). The software industry is establishing new models of music consumption with streaming music platforms and consequently creating development within the music industry.
The software that is created for content forwarding should not be considered only as a product of industry but as a mediator of the content (Manovich, Software Takes
Command 10). Consequently, I am using the field of software studies introduced by Lev
Manovich to examine the cloud music services as new media objects, i.e. a specific media channel meant for music. I explain his views later in this chapter. I am here connecting the
two fields of creative industries (the software and music industries); it is the combination of communication technologies and content that is creating a new medium. This places cloud music services into a group of media software that was defined and categorized by Lev Manovich (Software Takes Command 24).
Just as media needs content, so too does content also need media; this applies to media software as much as any other form of media or content. In the case of cloud music services and the evolution of software, a point has been reached where the software has begun to overtake the traditional means of music consumption (Manovich, Media after
Software 4). This is the first point about why we have to think about software as a creator of
popular music culture, as we previously had thought about the traditional radio or television (Gross, Gross, and Perebinossoff 21). Later in this thesis, I make the comparison between media software and traditional broadcasts more explicit; to do so I use Lynne Gross’, Brian Grossand’s and Philippe Perebinossoff’s book Programming for TV, Radio & The Internet. However, because the software is the primary concern in this thesis, we have to acknowledge that software is running on prewritten scripts that automate the entire process of data forwarding. In other words, the processes that are empowering the software are curating the user’s daily life, creating popular culture and indirectly affecting society. In this sense, cloud music services are becoming an automated music culture curator, which is proven in the later stages of the thesis.
One of the features responsible for this curation is the recommendation system, which offers a filtered set of results. In the software, we can find a number of different recommendation systems, from personalized filters to featured content filters, which are responsible for the curation of contemporary popular culture. The part of the software responsible for the recommendation tasks as well as for the curation of the moment are algorithms and, using the terminology of Alexander Galloway, what they produce is the algorithmic culture (Gaming Essays on Algorithmic Culture 4). One of the major goals of this thesis is to closely observe music curation algorithms within a music streaming software and to recognize the patterns of curatorship that occur as a result. Although there are currently a number of services offering music streaming and there is an extensive amount of newcomers, I chose to observe one of the currently most recognizable music streaming
1.2. BRIEF HISTORY
Music streaming services appeared as a result of digitalization. However, the digitalization was not solely the reason this software appeared; it was created as a consequence of disruptions in the music industry that happened because of digitalization.
Until 1999, the music industry had enjoyed the evolution of digital technology, as major labels started to sell music in digital formats, avoiding costs of physical media (Hracs 445). For the music industry, digital formats were seen as a tool of lowering the production costs and of increasing consumer prices, resulting in enormous profits as industry was creating its revenue from selling rights to the content it owned (Hracs 445). Although those digital formats seemed as a profitable technology in 1999, the music industry was faced with the so-‐called ‘MP3 Crisis’. The main issue had to do with the copyrighted material and the emergence of software formats that were easy to reproduce and share (Leyshon et al. 178). With the rapid development of personal computer technology and growing numbers of Internet users, online piracy also began to increase exponentially (Leyshon et al. 178). The music industry struggled with the crisis based on the traditional ways of dealing with copyright infringement, which did not result in positive customer relations (Makki). The music industry fought with lawsuits against software developers and users. One of the ideas of music industry was the introduction of the subscription model software that would allow users to access the music with the pay-‐per-‐song model, similar as in the early digital age (Hracs 449). One of the biggest music online resellers continues to successfully offer music based on this model: Apple’s iTunes Store.
Regardless of the music industry’s endeavours to decrease the amount of illegal content, its measures did not produce sufficient results (Hracs 449). However, independent and innovative developers introduced new models of online music consumption that attempted to endorse the interests of the music industry while simultaneously satisfying users’ expectations of free data exchange. Pressure coming from the music industry and from Internet users challenged developers to develop a concept that would resolve this crisis. On this, Spotify was built.
Spotify was designed from the ground up to combat piracy. Founded in Sweden, the home of The Pirate Bay, we believed that if we could build a service which was better than piracy, then we could convince people to stop illegal file-‐sharing, and start consuming music legally again. (“Spotify Explained”)
When Spotify was released, it announced that with this software the music industry is saved in terms of copyright regulation. Of course this was not the case, as Spotify alone could not
change the whole industry; however, due to its high recognition and successful model, it is appealing case for analysis.
1.3. DIGITAL MUSIC UBIQUITY
The ability to connect with the Internet has become a common practice in recent decades; however, using music cloud services to stream music at any time is a more recent technological development. The ability to stream music, not only on a smartphone, but also in the car, throughout the house and even during a taxi ride, brings the ubiquity of digital music to another level. This can be seen as another functionality that brought cloud music services closer to the traditional radio stations that were available for listening in every environment equipped with a transistor received. In this sense, digital technology was lacking the simplicity that was offered by analogue radio transistor, however with further digital development, music streaming offered new types of listening devices to be embedded into the selected environment.
In general, cloud music services allow their users to listen to music without transferring the files and offer them to stream music from various devices. This means that the cloud music services (and their software) have become ubiquitous. In software terminology, “ubiquity” means that the users can utilize the software at almost any time and any place; in this case, providing a nomadic digital experience (Niemelä and Latvakoski 71). To further develop this idea of ubiquitous streaming, we have to look into a more general discussion about music. Anahid Kassabian first introduced the term of “ubiquitous listening”, explaining that music (in comparison to most cultural products) is capable of being omnipresent in our lives and furthermore is consumed alongside or simultaneously with other activities (Kassabian 10). Software ubiquity combines computer technology that is used in everyday life, e.g. smartphone, personal computer, work computer, car, tablet, personal player, etc.; however, in this context we are talking about the continuous listening in everyday life contexts (Niemelä and Latvakoski 71). This is why it is necessary to distinguish between the ubiquitous software, which provides access to the software from all the places where devices are capable of connecting to the online network, and the ubiquitous listening, which is connected with the content, enabling us to add a background soundtrack to every moment of our lives. For the purposes of this thesis, I will use both terms to describe functionalities within the discussed software.
This terminology of ubiquitous listening and ubiquitous software is of great significance to this thesis as music cloud services are becoming increasing engaged in our lives through the omnipresent concept of music listening habits and are now trying to engage with users in as many situations as possible. The connection between ubiquitous listening and software is used to address the context of listening, according to Kassabian:
In general, music apps vary in relation to the level of activity they require, duration of interest they are likely to command, the degree of attention they may occupy, and so on. …it becomes clear, that it may well be productive to think of a group of iPhone apps as a cross between wearable and pervasive computing – on the one hand they are small and always with you, like wearable computing, and they can respond to your mood – but they both interact with and create your environment. […] IPhone apps are a new ‘size’ of interaction with environment, a new place of processing between wearable and pervasive computing, a new set of audio-‐visual relations, and a new form of soundscape management. (Kassabian 16)
Kassabian here explains how music has become even more involved in our life, since it is possible to access any kind of music with a device that has the access to it. Nevertheless, his theory is more connected with the description of music, saying that there are music and sounds that are capable to fit in every environment.
1.4. MUSIC STREAMING PLATFORMS ARE SOFTWARE
As I presented beforehand, cloud music platforms are online applications that offer music listening as a service. Furthermore, they are available on a number of different devices, which makes them ubiquitous, and they offer a wide range of content. In order to make the content interesting and convenient for the listener (also referred to as the “user” in this thesis), platforms offer a number of different techniques to organize and curate the data. As music-‐streaming platforms are software, content distribution within them is accomplished via the usage of pre-‐determined protocols, also known as algorithms. When I am referring to “algorithms”, I have in mind the explanation of Tarleton Gillespie, which presents algorithms as search engines of massive databases that manage our interactions on social networking sites and help us discover what is currently popular (Gillespie). He also discusses a particular type of algorithms, recommendation algorithms that are “suggesting new or forgotten bits of culture for us to encounter” (Gillespie). Software, as such, employs algorithms to provide user relevant content. As a result, software has already been a part of a vast number of debates on its influences on contemporary society. When I am writing about “society”, I refer to the 21st-‐century Western society and the nations influenced by it.
theoretical framework of software studies. He extended the theory of media studies by categorizing software and proclaimed it to be one of the influences of modern society (Manovich, Software Takes Command 20). In his book Software Takes Command, he argues that software currently is one of the engines of culture creation, and he termed this subset of software as a “cultural software” (Manovich, Software Takes Command 21). Furthermore, he presented another sublevel of cultural software, that he calls “media software”, which is being used for creating, editing, organizing, distributing, accessing and combining media content (Manovich, Software Takes Command 24). In this thesis, I look at the streaming music platforms as media software because they are enabling users to access music, record companies to distribute it, and, furthermore, they create capacity for classification and organization of songs.
As previously mentioned, I took the music-‐streaming platform, Spotify, as the object of study for this thesis. As a well-‐established software, that by June 10th 2015 had 20 million
subscribers and more than 75 million active users where available, it has created a digital environment that attracts new listeners (The Spotify Team). This environment is empowered by algorithms that are managing the content, linking music with users through specific channels within the software. Spotify is in this thesis as well described as well as a service, due to its delivery model; a subscription based software, known also by the name ”on-‐ demand software" The aim of this research is to focus on the recommendation algorithms that are responsible for music curation.
1.5. SOFTWARE AND ALGORITHMS
In order to discuss the topic of algorithms, I sketched an algorithm diagram (Figure 1), which I use through the text. As previously presented, a number of Spotify’s features are driven by algorithms. All algorithms employed in the software are a part of the first level of the algorithm diagram. Some of them are specifically employed to recommend music to users. These are the recommendation algorithms. They are a subgroup of all software algorithms, and thus are on the second level. There are a number of different recommendation algorithms currently available. In this thesis, however, I discuss two in particular. One type, which is already a part of many debates, is the personalization algorithm, most known among academics due to the work of Ali Pariser. He discusses personalization in depth within his book The Filter Bubble. I am contextualizing personalization with the help of Ali Pariser and Feuz Martin, Fuller Matthew, and Stalder
Felix, in their study of Google Search entitled Personal web searching in the age of semantic
capitalism: Diagnosing the mechanisms of personalization.
At the same level of personalization, I argue that there is another important recommendation algorithm that remains under-‐discussed. This the algorithm links users’ context with the software database in order to provide it with the most relevant matches. With the word “context”, I refer to the circumstances in which a person currently is. That can be a physical place or event, or a feeling or mood. To address this type of algorithm through this thesis, I use the term “context-‐curated algorithm” or “recommender”. The name is a combination of words context, where the circumstances are addressed, and curation as a process of organizing and displaying content. As a case study to present context-‐curated algorithms, I will use Spotify’s Featured Playlists, which can be seen on the fourth level. On this level of the algorithm diagram, there are some examples of algorithms currently in use. To address these algorithms. I am using Spotify’s Developer terminology found on their web page (“Spotify Developer”). Spotify’s developers have named the
algorithm responsible for personalized recommendations the “Discover” algorithm, and the one responsible for suggesting playlists the “Featured Playlists” algorithm.
As previously mentioned, I am focusing on Spotify’s Featured Playlists algorithm in this research, because it is linked not only with automated data forwarding but also with editorial suggestions. This is the reason we can find an external entity within the algorithm diagram (Figure 1). With “editorial suggestions”, I am referring to the music specialists who are responsible for contributions to music curation. Tarleton Gillepie recognized this difference between algorithmic and editorial logic in his article The Relevance of Algorithms (Gillespie). Based on what he calls ‘knowledge logic’:
Both struggle with, and claim to resolve, the fundamental problem of human knowledge: how to identify relevant information crucial to the public, through unavoidably human means, in such a way as to be free from human error, bias, or manipulation. Both the algorithmic and editorial approaches to knowledge are deeply important and deeply problematic (Gillespie).
I agree that both approaches are deeply important and deeply problematic, because a human can produce an error or a bias while algorithms are designed to automate human judgment (Gillespie). In the case of music streaming platforms, both approaches are being used, so this distinction is of great importance, even though some overlaps are expected.
If we return to the algorithm diagram, we can see that the recommendation algorithms are responsible for content forwarding (i.e. music forwarding in the case of Spotify). This means that these algorithms are mediating the cultural content. As a result, it could be argued algorithms are curating contemporary culture. Other scholars, specifically David Beer, Adrian Mackenzie, Tarleton Gillespie, etc., have already discussed digital culture curation; however, the case of Spotify has thus far remained understudied. For example, the well-‐known digital video platform Netflix have frequently been discussed among academics, due to its well-‐known recommendation algorithm (Gillespie 9; Beer Popular Culture and New
Media 36; Hallinan and Striphas 1). Furthermore, one of the most known examples in music
is the recommendation algorithm of the platform Last.fm, which has been discussed a number of times (Beer Popular Culture and New Media, 53; Beer Power through the
Algorithm?, 996). In the book Popular Culture and New Media, David Beer developed the
concept that recommendation algorithms can be understood as a way of shaping taste and of circulating the means of popular culture because they are suggesting what the users should pay attention to (86). This attention created by recommendations can be understood as curation and, because algorithms are responsible for taste shaping, this indicates the
power they hold. The power possessed by algorithms had been discussed by David Beer, when he presented how Last.fm’s recommendations enhanced post-‐hegemonic power (Beer, Power through the Algorithm? 997). His thesis was grounded on Scott Lash’s concept of “post-‐hegemonic power”, which argues that we currently live in the post-‐hegemonic era in which power is not coming from the institutions and their regime of representation, but it is influencing the society from the inside, in our case from software and algorithms (Lash 75).
1.6. RESEARCH QUESTION
In this thesis, I will present the algorithm behind Spotify’s Featured Playlists in depth and create a terminology to name and describe the processes that this kind of algorithm generates. Developing this theory will be my primary contribution to the field of software studies. With the use of empirical research, I will present how Featured Playlists are functioning and, as a secondary objective, I am will present a method of how to analyse music recommendations within Spotify. The results of this research will be relevant for algorithm researchers and for stakeholders of music streaming services: users, developers, artists and music industry representatives.
I will answer my primary question: What are the requirements to recognize an algorithm that belongs to the group of context-‐curated algorithms? Furthermore, I will answer the following question: What is the relation between algorithmic and editorial suggestions within the recommendation algorithms? As the final aim of the research, I will suggest how it would be possible to avoid the implications of authority within music recommendation systems.
1.7. OVERVIEW
In the introduction chapter of the thesis, I have presented my object of study. I have also introduced the theoretical framework to be used along with a brief history and presented the aims of the research. In the second chapter, I will introduce all major theories connected with the software and algorithms regarding how recommendation algorithms function. The following chapter will present the research methodology, specifically how the descriptive analysis and empirical analysis will be conducted. The fourth chapter will be focused on Spotify. I will present all research results of both descriptive and empirical
research and define suggested terminology to describe this kind of algorithmic behaviour. In the fifth chapter, I will summarize all research results and create a comparison of algorithmic and non-‐algorithmic curation with examples of present and past. The sixth chapter will be the conclusion of this thesis, where I will answer my research questions, present new possibilities for software curation and express my concerns regarding future software
2. L
OOKING INTO THE
S
OFTWARE
Currently, software is involved in many aspects of our daily life, and digitalization could be considered as significant as the invention of combustion engine or the harnessing of electricity (Manovich, Software Takes Command 8). It has a profound influence on society that needs to be examined. Here, I focus on music streaming services and examine one of the main features of Spotify platform. From the perspective of software studies, Spotify is a cultural software since it enables access to cultural artefacts (Manovich, Software Takes
Command 20). Furthermore, the service is also a part of a subset, media software
(Manovich, Software Takes Command 24). Spotify is one of many cultural programs used on a daily basis; in response to the increasing usage of such software, Manovich explained that:
[…] our contemporary society can be characterized as a software society and our culture can be justifiably called a software culture—because today software plays a central role in shaping both the material elements and many of the immaterial structures that together make up “culture.” (Manovich, Software Takes Command 33)
Software functions as an interface for users to use the data, but content management is done with the use of algorithms. In other words, algorithmic processes are ordering and sorting cultural content and deciding what is important for users (Beer, Popular Culture and
New Media 64; Manovich, Software Takes Command 33). The effects of algorithms that are
managing this kind of content have been described as algorithmic culture, firstly by Alexander Galloway and later by Ted Striphas (Galloway 18; Striphas). As a result, I am focusing on algorithms in this chapter in order to present how algorithms are functioning and what kind of influence they possess. In the second part of the chapter, I narrow my focus to selected recommendation algorithms in use by the music streaming service Spotify. However, first we need to understand how they function as a whole.
2.1. THE ERA OF ALGORITHMS
As I presented in the introduction, I am looking at algorithms on three levels. The first level is a wide one that combines different types of algorithms that are later more specified by their function. In general, they are functioning as software engines, and one could argue that they are functioning purely automatically. However, in some cases, they are subjected to the influence of external sources.
This has to be considered from two different angles. On one hand, we have to acknowledge the so-‐called “algorithmic objectivity” that is concerned with how algorithms are being developed and, on the other hand, the perspective in which algorithms already are functioning, but the input is being edited by human entities, in this thesis also referred to as “editors” (Gillespie). Looking at the algorithms from the user perspective, they can be seen as completely fair apparatuses “free from subjectivity or error” (Gillespie). However, Gillespie explains that this is not always a fact:
More than mere tools, algorithms are also stabilizers of trust, practical and symbolic assurances that their evaluations are fair and accurate, free from subjectivity, error, or attempted influence. But, though algorithms may appear to be automatic and untarnished by the interventions of their providers, this is a carefully crafted fiction. (Gillespie)
A similar theory was also created by Mackenzie, claiming that the primary issue within the algorithms is that the order that they create looks too natural and unmistakable (Mackenzie, 63). Just as editors of the traditional media have established the moral ethics of the content creation, Gillespie does so for algorithms. Here, the developers are addressed. Algorithmic objectivity is a part of a bigger debate about whether genuine objectivity is even possible; Evgeny Morozov claims that algorithmic objectivity, or with his terms, “neutrality” cannot ever be fully reached (145). To some extent, Nick Saver is also included in this debate in Knowing Algorithms in which he acknowledges the power of algorithms, saying that the solution to create algorithm objective is to make them transparent;
The solution is transparency: filters and the content they hide should be made visible (Savers 3).
This is a general statement, although recommendation algorithms of social music platforms can be used as an example. Because it is music they are curating, users would expect that the algorithms do this completely objectively; however, as I mentioned in the previous paragraph, that might not be the case.
Furthermore, we have to acknowledge another possible influence that affects the algorithmic output: editors. Content, as an input to an algorithm, can already be chosen by human entities, creating another level at which the output can be influenced. Gillespie also proposed this possibility; however, he stated that this can also be sometimes positive (Gillespie). I tend to agree that expert knowledge can be a welcome benefit to content curation, although the sense of authority cannot be avoided. Here, we can see a double influence that can be applied to the algorithmic mediation of (cultural) content. This means
that the outputs of the algorithms are generally subjected to a number of filtering layers before they reach the user.
The filtering processes within most algorithms are not transparent, and it could be argued that as a consequence algorithms possess influential power to its users. In general the proclamation of power within algorithms, has been acknowledged by Hamilton et al. and Scott Lash (Lash 75; Hamilton et al.). Here again we can take an example of recommendation algorithms for cloud music services. Recommendation algorithms possess power in a sense that they are suggesting music to their users and affecting their cultural taste (Beer, Power
through the Algorithm? 997). The power they possess is, however, not the same as power
over somebody, but in the way of shaping user experiences (Beer, Power through the
Algorithm? 997; Beer, Popular Culture and New Media 63).
In order to bypass this power within algorithms, we have to acknowledge what Saver suggested. His solution for lowering the power is by uncovering and making algorithms transparent (Savers 7). However, this might not be as easy as uncovering the code of their systems, since they are trade secrets (Savers 7). Here, I agree with Saver, a complete revelation of the code would be unacceptable for corporate entities. However, I am endorsing empowering users with the possibilities to edit the code. This would result in much lower levels of the power that we see now in algorithms.
The increasing importance of the power that is generated by the algorithms is not merely a discussion for digital technology researchers, media experts or developers. The power invested in music streaming platforms has also been recognized by the music industry. In the IFPI digital music report for 2014, they presented the statement of interest in the algorithmic recommendation within the cloud music services. For some time, the competition between them was based on the volume of music offered; however, this has now shifted to recommendations and music discovery (“Digital Music Report — IFPI — Representing the Recording Industry Worldwide”). Of course, both are connected to the music offered by platforms, but currently all of them are offering vast databases of content. This means that the competition between the services has shifted to quality of the
