View metadata, citation and similar papers at core.ac.uk brought to you by CORE New Media and Society provided by Repository of the Academy's Library

“Splendid Isolation”: The Reproduction of Geographic Inequality in Spotify’s Recommendation System

Journal: New Media and Society

Manuscript ID NMS-20-1205

Manuscript Type: Original Manuscript

extreme metal, , inequality, network analysis, popular music, Keywords: Forrecommendation Peer systems, Review Spotify, algorithms In this paper, we intend to tackle the issue of spatial inequality of music scenes on streaming platforms, through a network analysis case study of Hungarian extreme metal bands’ connections on Spotify, and by giving an overview of the role of algorithmic agents that shape and guide consumers’ access to musical content, and discovery of new music. We argue that the primary determinant of a given band’s international Abstract: connectedness in the ecosystem of Spotify is their international label connections. Those bands who are on international labels, have more reciprocal international connections, and are more likely to be recommended based on actual genre similarity. But bands who are signed with local labels or are self-published, tend to have rather domestic connections, and to be paired with other artists by Spotify’s algorithm according to their country of origin.

https://mc.manuscriptcentral.com/nms Page 1 of 20 New Media and Society

1 2 3 4 Introduction: Inequality in the digital music industry 5 As in various aspects of everyday life, in the operation of cultural industries, including the music industry, 6 7 the reproduction of social inequalities is a determinative factor. The unequal distribution of resources, 8 income and potentials can be discovered in the dimensions of ethnicity, nationality, geographical location 9 and gender, among others (Berkers and Schaap 2018). Inequalities created and reproduced in this field 10 affect all elements of the value chain of the music industry: they appear either in the spaces of production, 11 in the careers of musicians, creators or in the distribution networks and on the side of the consumers. 12 With the digitalization of the cultural industries, time to time resurfaced the idea that with the 13 ubiquitous presence of the internet, the music industry would become more level, equal and democratic, 14 so all those inequalities would slowly disappear (Hesmondhalgh 2019). As the advent of the internet was 15 accompanied by the expectation that the network will moderate offline inequalities, so were the first 16 17 music-centered social platforms generating similar hopes. The MySpace-boom was fueled partly by the 18 faith that the platform will open up unprecedented opportunities for everyone, regardless of location, 19 ethnicity or nationality. Similarly,For streaming Peer platforms Review have taken up the role of the revolutionary social 20 equalizer later. However, reviewing the data and analyses and meta-analyses regarding the current state 21 of the cultural industries, this desired condition has not been fulfilled. Instead of democratization and 22 equalization, rather the restructuring of power structures and hierarchies take place. In virtually all 23 segments of the digitalizing society (Eubanks, 2018), so too in the digital music industry, reproduction of 24 already existing social inequalities is still prevalent. Among other factors, female creators are still 25 26 underrepresented (Wang and Horvát, 2019), and artists operating from a central geographical location 27 have still way more opportunities to distribute and communicate their work (Verboord and Noord, 2016). 28 One of the most important fields of the reproduction of structural inequalities in the realm of 29 digital music is the selection and curation of the available musical catalog. The selection of musical content 30 by various methods and practices is prevalent since the advent of the music industry. Gatekeeper 31 decisions have determined in the pre-digital commerce the catalogues of brick and mortar stores 32 distributing items of recorded music (vinyl, cassettes), the representation of creators the broadcasting of 33 their works in media, including the live shows and the DJ selections, and those fundamental mechanisms 34 remained unchanged in the digital music industries. In the currently dominant streaming ecosystem still 35 36 gatekeeper decisions determine the selection and recommendation of musical content, only the nature 37 and agency of those gatekeepers have changed radically (Barzilai-Nahon, 2008; Wallace, 2018). The role 38 and place of the previously reigning big publishing companies have been taken by digital platforms 39 (Hesmondhalgh, 2019), that became new gatekeepers in all branches of the cultural industries by 40 aggregating content from various sources, and building on the wealth of data collected from their users’ 41 interactions and behavior (Vonderau, 2019). 42 In this paper, we wish to analyze how geographic inequalities are reproduced on the leading 43 digital music streaming platform by algorithmic agents that shape and guide consumers’ access to musical 44 45 content, and discovery of new music. We intend to demonstrate the workings of this process through a 46 network analysis case study. Our main question is: how spatial isolation of Central and Eastern European 47 (particularly Hungarian) extreme metal scenes is represented and reproduced by the “related artists” 48 feature of Spotify’s recommendation system? Throughout the case study and analysis we emphasize the 49 importance of reproduction, by arguing that patterns of unequal distribution of music consumption and 50 discovery are not novelties of the digital space, but rather have their roots in the already existing 51 inequalities of the music industry. 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/nms New Media and Society Page 2 of 20

1 2 3 Personalization and algorithmic recommendation systems in digital music 4 Communication of a musical catalogue, or a particular music towards the consumers in the physical space 5 6 of brick and mortar stores have been realized typically by guiding the focus of the customer gaze. 7 Recommendations, shelf barkers, sale signs tried to grab the attention of the consumer and to achieve 8 that they choose one particular item from the available hundreds or thousands. 9 In the musical ecosystem available on the digital visual interface – displays of different forms and 10 size – nevertheless, it is not possible anymore to present the catalogue and direct the attention of the 11 user in such ways. The range of visible options is fundamentally limited: even if it is possible to arrange 12 twenty to forty items on a larger screen, the same cannot be done on a small smartphone display, only a 13 couple of albums or songs can be shown, which is only a fraction of the options to be seen in a physical 14 store. Under such circumstances it is of paramount importance, what is included and what is not in this 15 16 limited selection, and if the user interacts with the interface (clicks, chooses, skips an item, makes a 17 purchase), what further options, paths are offered to them by the system. A particular paradox is created 18 in this context. We access the endless catalogue of the “infinite shelf”, as Anderson (2006) called it, though 19 an interface which offers usFor a way smallerPeer initial poolReview of choices than a brick and mortar store with 20 definitely finite shelf space. By the convergence of streaming services (Amazon Music, Apple Music, 21 Google Play Music, Spotify) platforms less and less aspire at selling individual music tracks to costumers, 22 but offer music as a service, as a „utility” (Goldschmitt and Seaver, 2019), and by doing this, they connect 23 the previously separated practices of music purchasing and listening. On a convergent platform, a user 24 25 can be a customer and a listener at the same time. 26 In navigating the vast digital catalogues, curatorial assistance is essential for the users (Jansson 27 and Hracs, 2018), and the same applies to the provider side. Although streaming platforms often employ 28 human and robotized curators as well, Spotify for instance in compiling playlists, and Pandora in analyzing 29 and coding musical, aural traits (Bonini and Gandini, 2019; Fry, 2019; Morris, 2012) managing the choices 30 and deemed preferences of millions of users with human operators exclusively is practically impossible, 31 so this task is outsourced to the non-human algorithmic agents (Ricci et al., 2015). This, naturally, does 32 not mean that algorithmic recommendation systems are purely robotic agents, as they are designed, 33 coded, run and maintained by human beings. Thus, the difference between human and non-human agents 34 35 should not be radicalized, as (Goldschmitt and Seaver, 2019): 65) put it “...these music discovery tools 36 should not be understood in terms of the popular opposition between people and algorithms, but rather 37 as sociotechnical systems that rely on and reinforce particular ideas about human and machine capacities 38 in relation to music.” 39 This process in the digital realm, in which human-designed, robotized algorithmic filtering and 40 recommendation systems play a central role, is called personalization altogether (Prey, 2018). At the core 41 of the conception of personalization is the idea that users, if possible, should encounter only contents 42 that correspond with their presumed taste and expectation the best. However, the name 43 44 “personalization” itself is rather misleading, as the goal of the platform strategy is not to serve all the 45 possible individual demands but, based on the available personal data – browsing history, previous user 46 behavior, social media connections and interactions and other contextual factors, such as check-ins – to 47 achieve a flawless, “frictionless” stream of content consumption (Prey, 2018; Rónai, 2020) which doesn’t 48 necessarily is the most diverse content possible (Snickars, 2017). In executing this strategy, algorithmic 49 recommendation systems get the main role. What is an algorithm, and what types of it are essential in 50 the digital music ecosystem? 51 52 53 Main types of algorithmic recommendation systems 54 Algorithmic recommendation systems are programs that depending on the nature of the incoming data 55 feed, make automated decisions according to the optimal outcome specified by their code, on various 56 levels of complexity (Li and Karahanna, 2015). The three most widespread types of robotized, algorithmic 57 58 59 60 https://mc.manuscriptcentral.com/nms Page 3 of 20 New Media and Society

1 2 3 filtering and recommendation systems are the item-based, content-based and collaborative filtering, and 4 the combination of these. Item-based recommendation systems make decisions based on the metadata 5 6 (the data on the data) of a given content item. For instance, whether they recommend or not a song 7 labeled as alternative country to a listener who previously had played music labelled with similar genre 8 tags. The same algorithm in case of a purchase transaction, might recommend another „similar” item, 9 which, according to the categorization system in the backend, is connected to the purchased one in one 10 or more ways. 11 Content-based filtering goes further in this regard, as monitors not only the metadata of the item, 12 but the actual content of it. If the acoustic traits of a given song are recorded and stored by the platform, 13 and recognized by algorithm, content-based filtering makes possible to recommend new songs and musics 14 based on pace, mood, aural factor, pitch, danceability or on even more detailed traits, such as the gender 15 16 of lead vocalist, uses of groove, level of distortion on the electric guitar, type of background vocals. Such 17 recommendation algorithms are used by, among others, one of the earliest entrants to the music 18 recommendation business, Pandora, based on their Music Genome Project (Fry, 2019; Morris, 2012; Prey, 19 2018), and The Echo Nest, whichFor is a partPeer of Spotify’s Review hybrid algorithmic recommendation system. (Prey, 20 2018). Probably the most ubiquitous filtering algorithm type, collaborative filtering, however, makes its 21 decisions based not on the traits on the songs themselves, but on the behavior of users, by monitoring 22 their past interactions on the given platform, and predicting their probable preferences (Prey, 2018). 23 These aforementioned algorithm types in most of the cases are not working alone, but compose 24 25 parts of a hybrid filtering/recommendation system’s decision-making mechanism, combined together. 26 Those hybrid recommendation systems may have several further functions or traits, such as creating a 27 user profile, based on the behavior and connections of the user. Furthermore, feedbacks from users can 28 be taken into account in various ways. Implicit feedback may be tracked and recorded: implicit feedback 29 is the type of feedback which does not require a dedicated action from the user, but their preferences can 30 be implied from the user behavior itself, such as skipping a song, listening songs in a particular order, 31 among others. Some personalization systems take into account explicit feedback-monitoring systems, 32 where users are given opportunities to express their opinion by directly evaluating particular features of 33 the service: for instance, writing reviews, giving stars, liking or disliking certain content. 34 35 36 Spaces of algorithmic reproduction: homogenization, bias, opacity 37 How are the operation mechanisms of algorithmic recommendation systems connected to the 38 reproduction of inequalities in the music industry? The various kinds of personalization systems have a 39 central role in the streaming ecosystem, in shaping listening to and purchasing and listening to music, and 40 also discovering new music (Aguiar and Waldfogel, 2018). In the wake of early initiatives such as Last.fm 41 or Pandora, today algorithms define the way how listeners, costumers have access to the vast catalogue 42 through the biggest music streaming platforms such as Spotify, Deezer, and as the largest online music 43 archive, YouTube-on is, and also the music player/seller outlets of platforms such as Amazon Prime Music, 44 45 Apple Music and Google Play Music. The algorithmic reproduction of inequalities may be realized in, 46 among others, three ways. The homogenization of choices, algorithmic bias, and algorithmic opacity, 47 whom are all deeply intertwined with the economic nature of the emerging ecosystem led by technology 48 giants using and operating the algorithms, the platform economy. 49 50 The homogenization of choices: from filter bubbles to frictionless music 51 One of the most popular buzzword in the popular, professional and academic discourse on digital media 52 53 is the “filter bubble” (also known as “echo chamber”), which was elevated into the public discourse by the 54 influential book of Eli Pariser (Pariser, 2011). According to the radical version of the filter bubble theory, 55 algorithms that monitor and analyze user behavior, map their presumed preferences, offer options to the 56 user that fit their presumed preferences – essentially the pattern of their past behavior – the best. Thereby 57 58 59 60 https://mc.manuscriptcentral.com/nms New Media and Society Page 4 of 20

1 2 3 a feedback loop is created from which it is no way to break out, users are not able to encounter with 4 opinions, perspectives, contents that differ significantly from their deemed taste or world view. However, 5 6 in their meta-analysis of the available academic literature on filter bubbles, Zuiderveen Borgesius and his 7 colleagues (Zuiderveen Borgesius et al., 2016) found that so far there is no solid empirical evidence on the 8 existence of filter bubbles per se and their total supremacy over our online content consumption patterns. 9 At the same time, it can be safely stated that if users rely exclusively on the feedback loop offered by the 10 algorithmic selection, their content consumption tends to become homogenized and polarized, in digital 11 music too (Fry, 2019). 12 For music streaming platforms, however, it is beneficial if the users do not intervene to the stream 13 of music pre-selected and offered by the algorithms, because the more automatic, seamless the music 14 consumption is, the more music is listened to by the user. (So the greater the revenue stream is for the 15 16 platform.) This is the notion what András Rónai calls the idea of frictionless music (Rónai, 2020), that 17 fundamentally defines the utopia of the platforms’ business models. The realization of the idea of 18 frictionless music facilitates and motivates such user behavior in which all factors (“frictions”) that might 19 stand in the way of continuous,For homogenous Peer music Review listening, are eliminated as much as possible. For 20 instance, on the user interface of Spotify, such is the function of the big green “shuffle play” button, 21 emphasized by the UX, and similarly, such is the goal of the various playlists based on different moods or 22 activities; to facilitate the continuous music listening. Viewed from the perspective of frictionless music, 23 traditional units of music consumption, such as albums or even genres, could be perceived as obstacles in 24 25 the reception of uninterrupted music stream offered by the algorithms. 26 The strategy and utopia of frictionless music is put in an even broader context by the transition of 27 digital interfaces from the visual to the voice user interface. Although market data is not transparent in 28 this regard, according to most of the estimates, in 2018 “tens of millions” were sold of the voice-operated, 29 so called home assistant devices globally. The market leader in this field is Amazon, with the software 30 Alexa (which can be found in various home appliances, including Amazon’s own home assistant device, 31 the Amazon Echo), and another important player is Google with the Google Assistant software and the 32 device Google Home (Tung, 2018). Expectedly, in a new digital ecosystem with a significant role of voice 33 user interfaces in it, personalization realized through algorithmic recommendation systems will have an 34 35 even more significant role in shaping music listening. As in the audio interface it is not possible to “show” 36 the listener more than one option at the same time, offering a constant stream of music might become 37 an even more important incentive for the providers. 38 39 Algorithmic bias 40 About algorithms, as invisible robots, codes, software, one might think that because of their rational, 41 mathematized construction, they lack all human traits, such as bias and erring. However, this assumption 42 43 does not hold true at least because of two reasons. First, as it was mentioned above, there are no 44 algorithms that operate fully without human intervention. The majority of digital platforms, thus the 45 music streaming services as well, use algorithms and human curation in an inseparable, intertwined way, 46 this is the reason why Bonini and Gandini (2019) calls the content filtering regime of the new platform 47 gatekeepers „algotorial power” (combined from the words algorithm and curator). The other reason is 48 that codes of algorithms are created by human beings, and the goals, ideas that are behind all the codes, 49 are invented by human agents too (Bozdag, 2013). 50 Algorithmic bias or unfairness (Speicher et al., 2018), beyond their definitely human context and 51 52 origin, can be detected in their automatized functioning too. One of these patterns through which bias 53 could manifest is the one which is behind the notion of frictionless music mentioned above. This is one of 54 the algorithmic spaces where human and non-human tendencies for partiality meet, as users tend to 55 select from a restricted circle of options that are not radically different from their previous choices; and 56 the algorithms continuously seeks to reinforce those decision making mechanisms. 57 58 59 60 https://mc.manuscriptcentral.com/nms Page 5 of 20 New Media and Society

1 2 3 One of the most common algorithmic bias patterns is the so called popularity bias. In this filtering 4 feedback loop, those items get in a better position, which are already in a better position. This tendency 5 6 is partly rooted in the inherent nature of collaborative filtering, as this type of filtering primarily weights 7 the interactions of users and not the content itself. As a result, a certain Matthew-effect takes place in 8 the recommendation system, putting aside the less popular creators and pushing forward even more 9 those who already reached a certain level of popularity (Bauer, 2019). Aside the popularity bias, an 10 algorithm can be partial in the selection of input data (selection bias) or in showing, framing the available 11 options to the users (presentation bias) (Bozdag, 2013). 12 If we look at the data and analyses available about and on Spotify, as the largest streaming 13 platform (almost) exclusively dedicated to music streaming, we can discover an array of biases – even 14 though it is still impossible to look into the codes of the recommendation system (Eriksson et al., 2019). 15 16 Thus the platform seems to reproduce the gender inequality already present in the music industry: among 17 the most streamed performers there are significantly more male artists than female ones – for instance, 18 in the end of 2017 in the top 10 chart of the most streamed songs, none of the performers were women 19 (Pelly, 2018). From an overview,For released Peer by Spotify Review in 2018, based on the data aggregated from the first 20 decade of the service, turned out that male performers dominated all the chart throughout the platform’s 21 ten year long history, and Anglo-Saxon artists had way more exposure (Snapes, 2018). 22 23 24 Algorithmic opacity 25 However, not directly but indirectly, the obscurity of the codes, and even values, policies and strategies 26 behind recommendation systems also contribute to the reproduction of inequalities (Ságvári 2017, Bodo 27 et al. 2017). As a result of this algorithmic opacity (also known as the “Black Box”), professional and 28 academic access to the operation of personalization systems is highly restricted, therefore, because 29 potential code audit methods (see Bodo et al., 2017) cannot be used to monitor and evaluate those 30 algorithms, in most of the cases the only remaining method to study the workings of algorithms is 31 deducting their probable guiding principles based on their output (often called as reverse engineering) – 32 33 in this paper we also chose this method, utilizing a smaller scale network analysis). All of this contributes 34 to the reproduction and maintenance of inequalities in at least two ways. On the one hand, by making 35 markets and competition intransparent: users, costumers do not have knowledge on the principles and 36 operations of the services they are dependent upon. On the other hand, denying access to the code makes 37 the audit and control of algorithms almost impossible (Eriksson et al., 2019). 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/nms New Media and Society Page 6 of 20

1 2 3 Case study 4 To understand better the workings and patterns of the reproduction of geographical inequalities in 5 6 algorithmic recommendation systems, we conducted an empirical case study based on the network 7 analysis of a sample of Hungarian extreme metal bands’ connections with other bands through Spotify’s 8 “related artist” feature. We complemented the network analysis with a qualitative examination of two 9 band’s individual networks. In the following, we will present the main research questions, the method of 10 11 data collecting, and finally, the results and discussion of the case study. The research questions were the 12 following: 13 14 Q1A: How do the Hungarian source bands’ outward connections (weighted by their position on the 15 "related artist" tab) to non-Hungarian bands correlate with the Hungarian source bands’ label background 16 17 (international v Hungarian) and language of lyrics? 18 Q1B: What is the country of origin of those non-Hungarian bands, to which the highest number of outward 19 ties goes from the HungarianFor source bands?Peer Review 20 Q2A: How do the Hungarian source bands’ reciprocal connections with non-Hungarian bands correlate 21 with the Hungarian source bands’ label background (international v Hungarian) and language of lyrics? 22 23 Q2B: What is the country of origin of those non-Hungarian bands, with which Hungarian source bands 24 have the highest number of reciprocal ties? 25 Q3: What are the main differences between the networks of two „typical” bands: one with international 26 reciprocal connections and one with only domestic connections? 27 28 29 Data and method 30 31 The sample of bands 32 The sample of source bands consists of 23 leading Hungarian extreme/modern metal bands, as significant, 33 important representatives of their respective subgenre (, , industrial, metalcore, 34 etc.). The bands were selected based on their position in the Hungarian scene: the majority of them 35 36 were/are on stage for more than a decade, and have a loyal regional, nationwide, or international follower 37 base, also their work is acknowledged by critics and fans as well. We included 20 currently active and 3 38 broken-up bands. The bands are on either Hungarian or international record labels. 39 40 The “related artists” page 41 Spotify’s recommendation system offers vast types of recommendation tabs based on different logics, but 42 they are all based on the connections between bands. The „related artists” tab (on desktop and in the 43 browser; the same feature is called “fans also like” in the android app for instance) is only one out of many 44 45 output versions of the recommendation system. Some of the other versions are the „More like...”, 46 „Because you listened to...”, „Similar to…”, „Suggested for you based on...” variations that can be seen 47 and found in different parts of the service. (The same recommendation system contributes to the 48 compilation of more complex, algotorial playlists, such as Discover Weekly or New Music Friday as well). 49 50 We selected the “related artist” tab because of two reasons. On the one hand, as opposed to the other 51 features, which differ from user to user (as each user sees their own “personalized” recommendation 52 feed), the “related artists” tab is “fixed”, and neither the featured bands, nor end users are capable of 53 changing its content. On the other hand, connections listed on the “related artists” tab are relatively 54 constant and stable over time. In a small-scale pilot study we conducted in February 2019, we recorded 55 56 the “related artists” on a smaller band sample, and compared them to the results of the latest data 57 58 59 60 https://mc.manuscriptcentral.com/nms Page 7 of 20 New Media and Society

1 2 3 collection. In the cases of the two particular bands whose network we analyze deeper in this paper, on 4 the related artist tab of the band Ektomorf, only one band changed out of 20, and for Apey and the Pea, 5 6 only 5 changed out of 20. Compared to these changes, other recommendation possibilities are more 7 volatile. 8 9 Data collection 10 The data collection was executed between 4 February and 9 February 2020. The reason for choosing this 11 period of the year was that traditionally it is one of the relatively uneventful periods in the music industry 12 – as opposed to, for example the summer festival season, or the release sales peak before Christmas. The 13 14 main method of data collection was desktop research. Regarding the attributes of bands, background 15 information and data were collected. Regarding the “related artist” tabs of the bands, we collected 16 information through the user interface. The result of this data collection are 23 edgelist tables, which also 17 include the “monthly listener” number and the countries of origin of all connected bands. We have also 18 collected all the relevant attributes related to the source bands of our sample. These attributes are the 19 For Peer Review 20 source bands’ monthly listeners on Spotify; data regarding other platform activities such as Facebook likes 21 and YouTube views; biographical data including years active, current status, English or Hungarian lyrics, 22 label of the latest album and that of previous one, and the country of origin of the publishers. 23 24 Structure of the data 25 The network is similar to a bipartite network, where one set of the nodes are the 23 Hungarian source 26 bands, and the other set of nodes consists of those bands who appeared on the “related artist” page of 27 28 the selected 23 Hungarian bands. However, it is not fully bipartite, as source bands could also appear on 29 other source band’s related artist page, so there could have been (and were) connections between the 30 first set of nodes too. Because of the logic of data collection, no connection was recorded between the 31 nodes of the second set. A one-way (outward) tie was recorded from the source band to another band, if 32 the other band appeared on the source band’s related artist page, but the source band didn’t appear on 33 34 the other band’s related artist page. A reciprocal tie was recorded between the source band and the other 35 band, if not only the other band appeared on the source band’s related artist page, but also the source 36 band appeared on the other band’s related artist page. Ties were weighted with the rank on the related 37 artist page with an integer between 1 to 20. Higher weights denote more importance. For instance, if a 38 band appeared first on the source band’s related artist page, the tie from the source band to that band 39 40 had a weight of 20. If a band was the last on the related artist page, the weight to that band was 1. The 41 same measure was applied in the calculation of the weights from non-source bands to source bands. 42 Figure 1 shows the same network with three different visualization. Panel A shows the 43 bipartiteness of the network, Panel B emphasize the clusters of ties going from the source bands to the 44 other bands, and Panel C illustrate that even the source bands connect to each other and that the non- 45 46 source bands are sometimes reached by several source bands. 47 48 49 50 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/nms New Media and Society Page 8 of 20

1 2 3 Figure 1: Three illustrations of the analyzed network 4 5 Panel A Panel B Panel C 6 7 8 9 10 11 12 13 14 15 For Peer Review 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 https://mc.manuscriptcentral.com/nms 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 Page 9 of 20 New Media and Society

1 2 3 Measures 4 Based on our research questions, the main attributes of the nodes (bands) were the following: label 5 6 background of the source band (Hungarian v international), language of lyrics of the source band, country 7 of origin of all bands. In determining if a source band was to be considered as signed with a Hungarian or 8 international record label, we looked at the publishers of their last (current) album’s publisher, and that 9 of released before the last one (penultimate). If at least one of the two last albums were published by an 10 11 international label, we recorded the band in question as having an international publishing background. 12 If none of their last two albums were published by international labels (or were self-published), we 13 considered them as having Hungarian publishing background. We classified the lyrics of the source bands 14 to the following categories: only Hungarian, only English, both. For this classification we took into account 15 all the lyrics of the band from the time the band was founded. 16 17 Results 18 In our first research question, we were focusing on the ties from the Hungarian source bands to the non- 19 For Peer Review 20 Hungarian bands: their number and importance. Table 1 shows the summary table of these connections. 21 The column out-degree denotes the number of ties going from the Hungarian source band to any non- 22 Hungarian bands. The number shows how many non-Hungarian bands appeared on the related artist page 23 of the given Hungarian source band. The average weights of out-degree were only calculated for those 24 source bands, who have at least one tie to non-Hungarian bands, so for those having at least one non- 25 26 Hungarian band on their related artist page. Weights were calculated as described above; thus, the 27 average weight shows the average importance of ties from the given source band to the non-Hungarian 28 bands. 29 Research question 1A was related to the correlation with the publisher and language of lyrics of 30 the Hungarian source bands. Those Hungarian source bands, who have Hungarian publishers connect to 31 32 1.9 non-Hungarian bands on average on their related artist page. While source bands with non-Hungarian 33 publishers have 17.3 non-Hungarian bands on average on their related artist page. If we also take into 34 account the importance of these connections, we can observe similar patterns. Hungarian source bands, 35 who have Hungarian publishers and at least one connection with non-Hungarian bands, also have least 36 37 important connections with these non-Hungarian bands. The average weight of these ties is 8.0 (on the 1 38 to 20 range). The same value at those Hungarian source bands who work with non-Hungarian publishers 39 is 9.8. 40 If we take into account the language of the lyrics of the Hungarian source bands, we can see 41 similar tendencies in the out-degree of the bands. Those Hungarian source bands, whose lyrics are only 42 43 in Hungarian have 10.0 connections to non-Hungarian bands on average. The same number is 12.9 at 44 those bands, whose lyrics are only in English – and interestingly even lower, 4.4 at those singing in both 45 languages. It suggests that those bands, who only sing in English have more non-Hungarian bands on their 46 related artist page compared to those singing only in Hungarian or in both languages. However, these 47 tendencies do not apply for the importance of the connections. If we only take into account those 48 49 Hungarian source bands, who have at least one connection to non-Hungarian (so have at least one non- 50 Hungarian band on their related artist page), we can observe that the average weight of the connections 51 is the highest (10.0) among those singing only in Hungarian, and only a little bit but lower (9.7) among 52 bands singing exclusively in English or in both languages (9.5). Both groups of bands have more important 53 connections on average compared to those bands, who sing both in English and Hungarian. These results 54 55 suggest that the language of the lyrics matters in the number of connections with non-Hungarian bands, 56 but not in the importance of these connections. However, it worth taking into account the ratio of bands 57 58 59 60 https://mc.manuscriptcentral.com/nms New Media and Society Page 10 of 20

1 2 3 having non-Hungarian connections in the group of those singing in Hungarian: it is 50% (2 out of the four 4 Hungarian singing band), while the same ratio in the group of those sing in English is 78% (7 out of the 9 5 6 bands). 7 8 9 Table 1: Summary table of ties from Hungarian source bands to non-Hungarian bands 10 11 Average weight of 12 Out-degree out-degree Publisher Language of lyrics 13 14 Aebsence 2.0 5.0 Hungarian Both 15 16 Blind Myself 0.0 Hungarian Both 17 18 Perihelion 17.0 10.0 Not Hungarian Both 19 Superbutt For0.0 Peer ReviewHungarian Both 20 21 Agregator 0.0 Hungarian Both 22 23 Christian Epidemic 0.0 Hungarian Both 24 25 Nevergreen 1.0 13.0 Hungarian Both 26 27 Ektomorf 20.0 10.0 Not Hungarian Both 28 29 AWS 0.0 Hungarian Both 30 31 Bridge To Solace 20.0 10.0 Not Hungarian English 32 33 Sear Bliss 20.0 10.0 Not Hungarian English 34 Bornholm 19.0 10.0 Not Hungarian English 35 36 The Southern Oracle 7.0 9.0 Not Hungarian English 37 38 Subscribe 0.0 Hungarian English 39 40 Wisdom 20.0 10.0 Not Hungarian English 41 42 Apey & the Pea 0.0 Hungarian English 43 44 Tormentor 20.0 10.0 Not Hungarian English 45 46 Harmed 10.0 9.0 Not Hungarian English 47 Dalriada 20.0 10.0 Hungarian Hungarian 48 49 Ørdøg 0.0 Hungarian Hungarian 50 51 Thy Catafalque 20.0 10.0 Not Hungarian Hungarian 52 53 Watch My Dying 0.0 Hungarian Hungarian 54 55 Pozvakowski 2.0 4.0 Hungarian Instrumental 56 57 58 59 60 https://mc.manuscriptcentral.com/nms Page 11 of 20 New Media and Society

1 2 3 4 5 Research question 1B was about the country of origin those non-Hungarian bands, that appeared the 6 most on the related artist page of the examined Hungarian bands. From all the 185 non-Hungarian bands 7 8 that were recorded in our network, the vast majority (173) appeared in only one Hungarian source band’s 9 related artist page. Eleven non-Hungarian bands appeared in two different Hungarian source bands’ 10 related artist page and there was only one non-Hungarian band that appeared three times. Table 2 shows 11 the name and country of origin of those bands appeared at least on two different related artist pages. The 12 column in-degree denotes the number of Hungarian source bands on whose related artist page the given 13 14 non-Hungarian band appeared. One can observe that the most common country of origin among these 15 bands is Norway, but we can find two bands from the UK in this list of 12 bands. The one band that 16 appeared in three different related artist page of Hungarian source bands is a UK band, called ‘Fen’. 17 18 Table 2: The name and country of origin of those bands, who appeared the most times on the Hungarian 19 source bands’ “related artist”For page Peer Review 20 21 In-degree Country 22 23 Helheim 2.0 Norway 24 25 Kampfar 2.0 Norway 26 27 Windir 2.0 Norway 28 29 Borknagar 2.0 Norway 30 Enslaved 2.0 Norway 31 32 Gwydion 2.0 Portugal 33 34 Negura Bunget 2.0 35 36 Rimfrost 2.0 Sweden 37 38 Winterfylleth 2.0 UK 39 40 Fen 3.0 UK 41 42 Lotus Thief 2.0 USA 43 White Ward 2.0 44 45 46 47 The second research question was related to the reciprocal ties between the Hungarian source bands and 48 the non-Hungarian bands. In this case – based on the definition above –, a reciprocated tie is present 49 50 between a Hungarian source band and a non-Hungarian band if the non-Hungarian band, which is present 51 on the related artist page of a given Hungarian source band also lists the Hungarian source band on its 52 related artist page. For instance if Ektomorf (a Hungarian source band) have Chimaira (a non-Hungarian 53 band) on its related artist page, and Chimaira also have Ektomorf on its related artist page, then the tie 54 between them is a reciprocated tie. Reciprocity is calculated by the number of reciprocated ties divided 55 56 by the number of all ties of a given band. In this case, for each Hungarian source band, reciprocity is 57 58 59 60 https://mc.manuscriptcentral.com/nms New Media and Society Page 12 of 20

1 2 3 calculated by the number of reciprocated ties with non-Hungarian bands divided by the number of all ties 4 with non-Hungarian bands. Thus, the indices of reciprocity can only be calculated if the Hungarian source 5 6 band has at least one tie to a non-Hungarian band. 7 Table 3 shows the summary of the characteristics of all Hungarian source bands regarding their 8 reciprocity value, the number of reciprocated ties, their publisher (Hungarian or not Hungarian) and the 9 language of their lyrics. Nine out of the twenty-three Hungarian source bands do not have any connections 10 11 with non-Hungarian bands based on their related artist page. Eight of the fourteen bands having at least 12 one non-Hungarian connection have no reciprocated ties. All in all, six bands had reciprocated ties with 13 non-Hungarian bands out of the examined twenty-three. Among those, where it was possible to calculate 14 (where the Hungarian source band had at least one tie to a non-Hungarian band), the average reciprocity 15 was 0.24 and the average number of reciprocated ties was 2.17. 16 17 Research question 2B focused on reciprocity and its correlation with the publisher and the 18 language of lyrics of the Hungarian source band. We emphasize that calculations could have been made 19 for those bands, which had atFor least one Peer tie to non-Hungarian Review bands. Those Hungarian source bands, which 20 have Hungarian publishers had the average reciprocity of 0.11 with non-Hungarian bands, while the same 21 number of those bands with non-Hungarian publisher was 0.29. If we focus on the number of reciprocated 22 23 ties between the source Hungarian band and the non-Hungarian bands, we can see even larger difference. 24 The average number of reciprocated ties is 0.46 among those with Hungarian publishers, and 4.4 among 25 those with non-Hungarian publishers. Both results suggest that a non-Hungarian publisher makes the 26 presence of reciprocated ties with non-Hungarian bands more likely. 27 If we take the language of the lyrics into account, we can observe contradictory tendencies. 28 29 Among those bands singing only in Hungarian, the average reciprocity is 0.62, while among those sing only 30 in English or in both languages, it is much less (0.18 and 0.20 respectively). The same trend is shown up in 31 the number of reciprocated ties. The average number of reciprocated ties is 4.75 in the group of Hungarian 32 singing source bands, but the same number is only 2.0 among those bands, which only have English lyrics 33 and 1.44 among those singing in both languages. For the understanding of these seemingly contradictory 34 35 results, it worth looking the number of source bands in the different groups. There are only four source 36 bands, which only sing in Hungarian and out of these, two of them have non-Hungarian connections and 37 these two bands are both reciprocal connections too. At the same time, there are nine bands, which only 38 have English lyrics, from which all the nine bands have at least one connection with a non-Hungarian band. 39 From these nine bands, three of those have reciprocal connections. Thus, the bases of the calculations – 40 41 namely the number of those bands, who have at least one tie to a non-Hungarian band – are different in 42 the two groups. 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/nms Page 13 of 20 New Media and Society

1 2 3 Table 3: Summary table of reciprocated ties between Hungarian source bands and non-Hungarian bands 4 5 N of reciprocated Language of 6 Reciprocity ties Publisher lyrics 7 8 Aebsence 0.0 0.0 Hungarian Both 9 10 Blind Myself 0.0 Hungarian Both 11 12 Superbutt 0.0 Hungarian Both 13 14 Agregator 0.0 Hungarian Both 15 16 Christian Epidemic 0.0 Hungarian Both 17 18 Nevergreen 0.0 0.0 Hungarian Both 19 For Peer Review 20 AWS 0.0 Hungarian Both 21 Ektomorf 0.79 13.0 Not Hungarian Both 22 23 Perihelion 0.0 0.0 Not Hungarian Both 24 25 Apey & the Pea 0.0 Hungarian English 26 27 Subscribe 0.0 Hungarian English 28 29 Sear Bliss 0.0 0.0 Not Hungarian English 30 31 Bornholm 0.1 1.0 Not Hungarian English 32 33 Bridge To Solace 0.52 7.0 Not Hungarian English 34 35 Harmed 0.0 0.0 Not Hungarian English 36 Tormentor 0.67 10.0 Not Hungarian English 37 38 The Southern Oracle 0.0 0.0 Not Hungarian English 39 40 Wisdom 0.0 0.0 Not Hungarian English 41 42 Watch My Dying 0.0 Hungarian Hungarian 43 44 Ørdøg 0.0 Hungarian Hungarian 45 46 Dalriada 0.46 6.0 Hungarian Hungarian 47 48 Thy Catafalque 0.79 13.0 Not Hungarian Hungarian 49 Pozvakowski 0.0 0.0 Hungarian Instrumental 50 51 52 53 Altogether we have found 50 bands, which had reciprocated connections with at least one Hungarian 54 source band. As Table 4 shows, the relative majority (11) of these bands are from the USA, 8 are from 55 56 57 58 59 60 https://mc.manuscriptcentral.com/nms New Media and Society Page 14 of 20

1 2 3 Norway and 7 are from Germany, and 3-3 are from Finland and Denmark. There are maximum two bands 4 from other countries, which have reciprocated connections with the Hungarian source bands. 5 6 Table 4: The number of non-Hungarian bands with reciprocated ties with the Hungarian source bands 7 8 by country 9 Country N of bands 10 11 USA 11 12 13 Norway 8 14 15 Germany 7 16 17 Finland 3 18 19 Denmark For3 Peer Review 20 21 Italy 2 22 23 Greece 2 24 2 25 26 Sweden 2 27 28 Brazil 2 29 30 Czech 31 Republic 1 32 33 Japan 1 34 35 Iceland 1 36 37 France 1 38 39 Australia 1 40 41 Israel 1 42 43 Switzerland 1 44 Brasil 1 45 46 47 48 The third research question was related to two extreme cases of source bands. These cases are presented 49 by two case-studies about two bands. The first selected Hungarian source band is called Apey and the Pea 50 51 (who recently changed their name to Lazarvs) only having Hungarian connections. The second band 52 selected is called Ektomorf and it has only non-Hungarian connections. Figure 2 shows the network of 53 these two Hungarian bands. The figure denotes several characteristics of the network. The middle node 54 is the selected Hungarian source band. The closer is another band to the source band, the stronger the 55 relationship between them (according to the weight of the tie from the source band). Green ties denote 56 57 58 59 60 https://mc.manuscriptcentral.com/nms Page 15 of 20 New Media and Society

1 2 3 reciprocal connections, black one one-direction connections from the source Hungarian band to the other 4 band. The width of green ties is proportional to the weight of the tie from the band to the source band. 5 6 The size of the node of a band is proportional to the number of monthly listeners of the band on Spotify. 7 8 9 10 11 12 13 14 15 16 17 18 19 For Peer Review 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/nms New Media and Society Page 16 of 20

1 2 3 Figure 2: The network of two extreme patterned Hungarian source band: Apey and the Pea and Ektomorf 4 5 6 7 Panel A: Apey and the Pea Panel B: Ektomorf 8 9 10 11 12 13 14 15 For Peer Review 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 https://mc.manuscriptcentral.com/nms 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 Page 17 of 20 New Media and Society

1 2 3 Discussion 4 Geographical inequality, as a relational concept, can take many forms: each field of cultural production 5 6 and consumption has their own centers and peripheries. We chose to analyze this particular segment of 7 the popular music field, as in a specific context patterns of geographical inequalities can be even more 8 apparent. In the case of extreme metal, traditionally the United States, Germany, the UK, and three of the 9 Nordic countries – Finland, Norway and Sweden – are considered to be the global centers (DeHart, 2018; 10 Maguire, 2015) and this unequivocally shows in the recommendation system as well. Regarding research 11 questions 1 and 2, the results show that if a Hungarian band has an international label background, it 12 significantly elevates their chances to be recommended by a non-Hungarian band’s “related artists” tab. 13 Moreover, if a connection is formed, it is very likely that that the connected bands are from the global 14 15 centers of the genre (9 out of 12 in the case of outward connections [Table 2], and 31 out of 50 bands in 16 the case of reciprocal connections [Table 4].) What is even more striking, there are almost no ties between 17 countries of the CEE region: local bands are most likely isolated, but if not, then they are rather connected 18 to the global centers directly (we found 2 regional out of 12 most frequent outward connections, and 1(!) 19 out of 50 regional reciprocalFor connections), Peer as it can Review be seen in the offline regional dynamics (Author, 20 2012). 21 Another important takeaway from the data is that bands who have international reciprocal 22 connections, are recommended based on actual genre similarity more likely, but bands with only domestic 23 connections tend to be paired by Spotify’s algorithm according to their country of origin. As it can be seen 24 25 by comparing the networks of the bands Apey and the Pea and Ektomorf (regarding research question 3), 26 the stronger the international relations are, the more genre-based are the recommendations. For 27 example, Apey and the Pea have only Hungarian bands on their “related artists” page, but the listed bands 28 are very heterogeneous, representing various subgenres, the only common in the enlisted bands is that 29 they are all Hungarian. As a contrast, Ektomorf’s international connections are very genre-specific, and 30 even more importantly, they have a very strong presence on the “related artists” pages of leading bands 31 of the genre. Another important difference between the Hungarian-only and the international network is 32 the number of potential users who might click through to visit the source band’s page. In the case of Apey 33 34 and the Pea, the sum of the monthly listeners of the bands they have reciprocal connections with is 35 11,037, with and average of 920. The same values in the network of Ektomorf are 3,059,658 and 235,358, 36 meaning Ektomorf has so much bigger of a network involving a 255 times larger audience to potentially 37 interact with their music. 38 39 Conclusions 40 In this paper, we intended to tackle the issue of algorithmic inequality fostered by streaming platforms, 41 through a network analysis case study. Unequal distribution patterns of music consumption and discovery 42 did not disappear with the advent of the algorithm-driven digital space, but rather have reappeared and 43 44 got reformulated in the digital cultural industries (Hesmondhalgh, 2019). For instance, in the realm of 45 digital music, among other factors, female creators are still underrepresented (Wang and Horvát, 2019), 46 and artists, bands operating from a central geographical location still have way more opportunities to 47 distribute and communicate their work (Verboord and Noord, 2016). Under the auspices of the ubiquitous 48 algorithmic personalization efforts (Prey, 2018), controlled by platforms acting as the new gatekeepers 49 (Aguiar and Waldfogel, 2018), practicing their newfound ‘algotorial power’ (Bonini and Gandini, 2019) 50 algorithmic bias and inequality (Bauer, 2019; Goldschmitt & Seaver, 2019) fosters unequal music 51 consumption and discovery patterns in the streaming ecosystem. 52 53 How could such patterns of inequality be reproduced by algorithms in a particular music scene? 54 In order to approach the phenomenon, we asked: how geographic isolation of Hungarian extreme metal 55 bands might be represented by the “related artists” feature of Spotify’s recommendation system? With 56 the tools of network analysis, combined with qualitative methods, we mapped out the connections 57 58 59 60 https://mc.manuscriptcentral.com/nms New Media and Society Page 18 of 20

1 2 3 between a sample of selected bands (n=23) and those bands that were featured on their “related artists” 4 page (n=20 per each band). We distinguished three levels of connectedness: on level one no international 5 6 connections were suggested on the “related artists” page, on level two there were one or more 7 international connections, on level three one or more of the international connections were reciprocal – 8 meaning that the “source” band was featured on the related artist’s “related artist” page too. 9 From the results, it can be seen that the way bands are represented in the recommendation 10 system – or, in other words, the way they are connected to other bands – significantly overlaps with their 11 offline connections in the music industry. Those bands who are signed with international labels, have 12 more level three connections, and are more likely to be recommended based on genre similarity. But 13 bands who are published by Hungarian labels or are self-published, tend to have level one connections, 14 and tend to be paired with other artists by Spotify’s algorithm according to their country of origin. Also, 15 16 the stronger the international connections are, the more genre-based are the recommendations. Based 17 on that sample it seems like the primary determinant of the nature of outward and reciprocal connections 18 in the recommendation system is the nature of label connections, this way the algorithm represents and 19 reproduces the bands’ geographicalFor (dis)advantagePeer Reviewat the same time. However, we emphasize that not 20 only algorithms are to blame for reproducing already observable inequalities in the music industry. 21 Besides the ultimate human design of algorithms and various human interventions in their functioning, 22 human listeners signal and share their decisions via the interface while interacting with algorithms. By 23 unearthing such invisible patterns of a particular recommendation system, we rather aimed at a better 24 25 understanding of how the socio-technical system of personalization works and reproduces existing 26 inequalities in the streaming ecosystem. 27 28 29 References 30 Author (2012) 31 Aguiar L and Waldfogel J (2018) Platforms, Promotion, and Product Discovery: Evidence from Spotify 32 Playlists. JRC Digital Economy Working Paper; JRC Technical Reports, JRC112023. 33 Anderson C (2006) The Long Tail: Why the Future of Business Is Selling Less of More. Hyperion. 34 Barzilai-Nahon K (2008) Toward a Theory of Network Gatekeeping: A Framework for Exploring 35 36 Information Control. Journal of the American Society for Information Science and Technology 59(9): 37 1493–1512. 38 Bauer C (2019) Allowing for equal opportunities for artists in music recommendation. Proceedings of the 39 1st Workshop on Designing Human-Centric MIR Systems (wsHCMIR 2019), satellite event to 20th 40 annual conference of the International Society for Music Information Retrieval (ISMIR 2019): 25–27. 41 Berkers P and Schaap J (2018) Gender Inequality in Metal Music Production. Emerald Publishing Limited. 42 Bodo B, Helberger N, Zuiderveen BF, et al. (2017) Tackling the Algorithmic Control Crisis – the Technical, 43 Legal, and Ethical Challenges of Research into Algorithmic Agents. The Yale Journal of Law & 44 45 Technology 19(1): 133–180. 46 Bonini T and Gandini A (2019) “First Week Is Editorial, Second Week Is Algorithmic”: Platform 47 Gatekeepers and the Platformization of Music Curation. Social Media and Society 5(4): 1–11. DOI: 48 10.1177/2056305119880006. 49 Bozdag E (2013) Bias in Algorithmic Filtering and Personalization. Ethics and Information Technology 50 15(3): 209–227. DOI: 10.1007/s10676-013-9321-6. 51 DeHart C (2018) Metal by numbers: Revisiting the uneven distribution of . Metal 52 Music Studies 4(3): 559–571. DOI: doi: 10.1386/ mms.4.3.559_1. 53 54 Eriksson M, Fleischer R, Johansson A, et al. (2019) Spotify Teardown : Inside the Black Box of Streaming 55 Music. MIT Press. 56 Eubanks V (2018) Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. 57 58 59 60 https://mc.manuscriptcentral.com/nms Page 19 of 20 New Media and Society

1 2 3 Martin’s Press. 4 Fry A (2019) An Echoic Chamber: Algorithmic Curation and Personalised Listening. In: Mazierska E, Gillon 5 6 L, and Rigg T (eds) Popular Music in the Post-Digital Age Politics, Economy, Culture and Technology. 7 Bloomsbury Publishing, pp. 269–290. 8 Goldschmitt KE and Seaver N (2019) Shaping the Stream: Techniques and Troubles of Algorithmic 9 Recommendation. In: Cook N, Ingalls MM, and Trippett D (eds) The Cambridge Companion to 10 Music in Digital Culture. Cambridge University Press, pp. 63–81. DOI: 10.1017/9781316676639.006. 11 Hesmondhalgh D (2019) Have digital communication technologies democratized the media industries? 12 In: Curran J and Hesmondhalgh D (eds) Media and Society, 6th Edition. London, New York: 13 Bloomsbury Academic, pp. 101–120. 14 Jansson J and Hracs BJ (2018) Conceptualizing Curation in the Age of Abundance: The Case of Recorded 15 16 M usic. Environment and Planning A 50(8): 1602–1625. DOI: 10.1177/0308518X18777497. 17 Li SS and Karahanna E (2015) Recommendation Systems in a B2C e-Commerce Context: A Review and 18 Future Directions. Journal of the Association for Information Systems 16(2): 72–107. 19 Maguire D (2015) DeterminantsFor of the Peer production ofReview heavy metal music. Metal Music Studies 1(1): 155– 20 169. DOI: 10.1386/mms.1.1.155_1. 21 Morris JW (2012) Making music behave: Metadata and the digital music commodity. New Media & 22 Society 14(5): 850–866. DOI: 10.1177/1461444811430645. 23 Pariser E (2011) The Filter Bubble: What the Internet Is Hiding from You. The Penguin Press. 24 25 Pelly L (2018) Discover Weakly: Sexism on Spotify. The Baffler, 4 June. 26 Prey R (2018) Nothing personal: algorithmic individuation on music streaming platforms. Media, Culture 27 and Society 40(7): 1086–1100. DOI: 10.1177/0163443717745147. 28 Ricci F, Rokach L and Shapira B (eds) (2015) Recommender Systems Handbook. Second edi. Springer. 29 DOI: 10.1016/0002-9343(62)90176-6. 30 Rónai A (2020) Frictionless Platforms, Frictionless Music: The Utopia of Streaming in Music Industry 31 Press Narratives. In: Tofalvy T and Barna E (eds) Popular Music, Technology, and the Changing 32 Media Ecosystem. Palgrave Macmillan, pp. 97–113. DOI: https://doi.org/10.1007/978-3-030- 33 44659-8_6. 34 35 Ságvári B (2017) Diszkrimináció, átláthatóság és ellenőrizhetőség: bevezetés az algoritmusetikába. 36 Replika 27(103): 61–79. 37 Snapes L (2018) Spotify at 10: men dominate streaming service’s most-played artists. The Guardian, 10 38 October. 39 Snickars P (2017) More of the Same – On Spotify Radio. Culture Unbound. Journal of CurrentCultural 40 Research 9(2): 184–211. DOI: https://doi.org/10.3384/cu.2000.1525.1792. 41 Speicher T, Heidari H, Grgic-Hlaca N, et al. (2018) A unified approach to quantifying algorithmic 42 unfairness: Measuring individual & group unfairness via inequality indices. Proceedings of the ACM 43 44 SIGKDD International Conference on Knowledge Discovery and Data Mining: 2239–2248. DOI: 45 10.1145/3219819.3220046. 46 Tung L (2018) Amazon: We sold tens of millions of Echo devices in 2018, and Alexa now has 70,000 skills. 47 ZDNet, 20 December. 48 Verboord M and Noord S van (2016) The online place of popular music: Exploring the impact of 49 geography and social media on pop artists’ mainstream media attention. Popular Communication 50 14(2). Taylor & Francis: 59–72. DOI: 10.1080/15405702.2015.1019073. 51 Vonderau P (2019) The Spotify Effect: Digital Distribution and Financial Growth. Television and New 52 53 Media 20(1): 3–19. DOI: 10.1177/1527476417741200. 54 Wallace J (2018) Modelling Contemporary Gatekeeping: The rise of individuals, algorithms and platforms 55 in digital news dissemination. Digital Journalism 6(3). Routledge: 274–293. DOI: 56 10.1080/21670811.2017.1343648. 57 58 59 60 https://mc.manuscriptcentral.com/nms New Media and Society Page 20 of 20

1 2 3 Wang Y and Horvát EÁ (2019) Gender differences in the global music industry: Evidence from 4 MusicBrainz and the Echo Nest. Proceedings of the 13th International Conference on Web and 5 6 Social Media, ICWSM 2019 (Icwsm): 517–526. 7 Zuiderveen Borgesius FJ, Trilling D, Möller J, et al. (2016) Should we worry about filter bubbles? Internet 8 Policy Review 5(1): 1–16. DOI: 10.14763/2016.1.401. 9 10 11 12 13 14 15 16 17 18 19 For Peer Review 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/nms