Sage Journals: Discover world-class research

Abstract

‘Diversity’ is a heavily freighted and multivalent keyword in the global digital media environment. The recommender systems used by platforms are particularly acute sites of development and debate around the political, cultural and technical issues ‘diversity’ signifies. Drawing on a review of computer science publications on recommender systems in media and entertainment as well as a survey of recent advances in media and cultural policy scholarship, this short article performs a pragmatic close reading of diversity in these intersecting fields. We note that attention must be paid to the specific challenges and politics of diversity not only in particular cultural fields but also in local cultural contexts, drawing on examples from music and SVOD platforms to flesh out these questions and the practical possibilities that arise from them.

Keywords

cultural diversity cultural policy digital media platforms recommendation recommender systems

Introduction

The recommender systems that drive platform features like TikTok's ‘For You Page’, YouTube's ‘up next’ feature, and Spotify's ‘Discover Weekly’ playlists have become central to ordering content, driving user engagement, and maximising advertising value for digital media and entertainment platforms. They are multi-stakeholder systems that mediate multi-sided markets, appealing simultaneously to the end users receiving recommendations, the content creators or media producers reliant on algorithmic distribution, and the advertisers seeking to reach relevant audiences.

In recommending content to users, global platforms must also navigate the national and international laws and business rules that impact what content is available to which audiences. These include market segmentation wrought by copyright licensing, local content requirements in those jurisdictions that have extended this regulation to subscription video on demand services (SVODs), for example, and content restrictions imposed through classification-style regimes, morality laws, or ‘brand safety’, which can impact materially on the careers of online creators. And, in an environment of increasing public concern about the impact of ‘algorithms’ on society, platform companies also have an interest in convincing regulators, advertisers and other third parties that they are curating content or items responsibly (Helberger et al., 2018).

The idea of diversity shows up in relation to all these entangled and competing interests – interests which recommender systems need to negotiate and balance. In this short article, we sketch some of the competing meanings attached to ‘diversity’ in the context of various media and cultural platforms’ recommender systems and by a range of humanities, social science and technical disciplines, and suggest some of the areas where these different meanings offer potential for collaboration in the public interest.

Diversity in recommender system design

Recommender systems are ‘prediction machines’ (Agrawal et al., 2018): they use data and machine learning to suggest content the system bets users will engage with. Platforms present these suggestions to users via features such as newsfeeds, custom playlists, and personalised landing pages. To make their predictions, recommender systems generally combine two basic principles: ‘collaborative filtering’ and ‘content-based’ recommendation.

Collaborative filtering draws on patterns of user practices across a section of the platform's population. To take music streaming as an example, suppose Spotify users who frequently listen to Lady Gaga also tend to listen to Beyoncé. In that case, it makes sense to recommend Beyoncé to those Lady Gaga fans who have not yet discovered Beyoncé. ‘Content-based’ recommendation is based on an individual listener's engagement with diverse types of musical content: if a user frequently listened to 1960s jazz during the past week, it is reasonable to recommend more jazz from the 1960s to this user during the week ahead.

Traditionally, when recommender system designers evaluate the effectiveness of a recommender system, they focus on the accuracy of its predictions – that is, the likelihood that, presented with an item of content, the user will engage with it. However, in practice, the more accurate a recommender system is, the more it risks inducing monotony, or ‘reduced consumption diversity’ (Anderson et al., 2020) – and therefore user disengagement, which runs counter to the business logics of platforms. Hence, recommender system designers already try to balance the accuracy measure with measures of diversity (e.g. Kunaver and Požrl, 2017). Even when understood very simply, however, balancing these two measures is difficult. If a recommender system is too accurate, an ‘overfitting’ problem can occur, whereby content in the recommendation chain is so similar that it is redundant (de Gemmis et al., 2015); on the other hand, recommendations intended to increase the diversity of a user's experience run the risk of simply being too far away from what the user actually wants, risking disengagement. This balancing act is commonly referred to as the ‘diversity-accuracy dilemma’ (Li et al., 2020).

Within the branch of computer science scholarship that underpins recommender system design and evaluation, diversity is simply a measure of the number of distinct items (songs, videos, news articles) that appear in a list of recommendations (Kunaver and Požrl, 2017). However, the differences or similarities among these items can be measured using a number of dimensions, introducing greater complexity and nuance – and it is here that there is potential for enrichment and experimentation, as well as the need for medium-specificity.

On music streaming platforms, for example, diversity can be measured through collaborative filtering that computes recommendations based on user data input; content-based approaches that index musical metadata and annotation added by humans; and contextual approaches, such as environmental factors (time of day, weather, holiday events); or psychological factors (personality, mood, social setting) (Schedl et al., 2015). Even more possibilities exist when computationally analysing musical audio files, including features such as those Spotify labels as ‘danceability’, ‘acousticness’, ‘energy’, ‘instrumentalness’, ‘liveness,’ ‘loudness’, ‘speechiness’, ‘tempo’, and ‘valence’ (Spotify, n.d.). We would argue such measures could be reconnected to cultural concepts relating to aesthetics, genres, and subcultures, suggesting creative possibilities for collaboration with computer scientists.

There is plenty to build on: cultural sociologists have a long tradition of studying different aspects of diversity in cultural fields such as popular music for reasons concerning questions of taste and artist welfare, and their approaches could be adapted for use in the contemporary era. Starting in the 1970s, sociologists (e.g. Lopes, 1992; Peterson and Berger, 1971) examined how aggregate consumption diversity, operationalised as the number of unique songs that reached a certain level of popularity, evolved over time. Other studies focused on content diversity in terms of genre or acoustics (e.g. Bourreau et al., 2022). Meanwhile, some economists (e.g. Aguiar and Waldfogel, 2021) addressed geographic aspects of diversity, such as the distribution of commercially successful artists across countries. The questions driving these studies were primarily focused on how diversity was impacted by factors such as industry concentration (Peterson and Berger, 1971) or digitisation (Bourreau et al., 2022) and explain why diversity fluctuated over time.

While we have so far mainly discussed music systems, there is much experimentation and variety across other kinds of recommender systems, for various media forms, further complicating the challenge of defining and evaluating diversity (Morales et al., 2021).

Representational dimensions of diversity

In addition to medium-specificity, a cultural studies approach to analysing and improving diversity in recommender systems would incorporate moral questions such as fairness, justice, and equality. Dimensions of recommendation addressing these questions might include the number and prominence of content creators (musicians, authors, or YouTube creators) representing a range of identity backgrounds; the amount and prominence of content representing a range of viewpoints or perspectives; or approaches designed to surface ‘niche’ or ‘long tail’ content, satisfying users with less mainstream preferences and avoiding popularity bias. Some advances in this respect have been made in work focused on news (Hilden, 2022); scholars and policy-makers are testing long-standing questions of media diversity (Napoli, 1999) against the operations of public service media (PSM) systems, and using their findings as the basis of policy and technical innovations (Helberger, 2018).

Scholarship on social media entertainment and the creator economy also highlights issues of inequity and bias in recommender systems tied to visibility of particular social identities and bodies (see Duffy and Meisner, 2023; Gillespie, 2022; Nicholas, 2022). Content featuring women of colour, trans people, people with disability, and plus-sized bodies are among the types of content that creators report being less likely to benefit from – or even to be suppressed by – algorithmic recommendation (Middlebrook, 2020), although platforms make it remarkably difficult for creators to prove these claims (Cotter, 2021).

Beyond concerns about unequally distributed attention, there are also issues of harmful representation, stereotyping and the commodification of diverse identity groups through recommendation. Noble's (2018) work was pioneering in this respect, finding that commercial search engines returned results that reinforced the historic sexualisation and dehumanisation of Black girls. Gerrard and Thornham's (2020) analysis of Pinterest's recommender system revealed the ways it connects content about eating disorders with ideals of ‘consumerist, youthful and white’ femininity, arguing that ‘what gets generated through recommendation systems are over-simplified versions of gender and other identities’ (2020: 1279) that tie in with the commercial objectives of platforms and advertisers.

Geographic dimensions of diversity

Geographic dimensions of diversity are important to consider in the context of global digital platforms (Steinberg and Li, 2017). To take the example of streaming video on demand (SVOD) platforms, whose politics link to long-standing debates about media globalisation and related policy concerns, part of the strategic advantage of platforms like Netflix is their ability to ‘imagine [and algorithmically target] their subscribers as transnational clusters of tastes and sensibilities’ (Lotz, 2021: 207; emphasis added): ‘[b]y appealing to discourses of global citizenship and diversity, Netflix brands itself as translator across cultures able to speak to everyone’ (Asmar et al., 2023: 36). Indeed, Netflix engineers claim to have transitioned to a ‘global’ (rather than region-specific) recommendation algorithmic system in 2016 in order to capitalise on user preference data at scale and in part to compensate for the company's lack of data in countries with comparatively few subscribers (Gomez-Uribe, 2016).

In reality, global platforms remain locally situated, albeit in uneven ways. They must negotiate (or bend to) state power in a variety of jurisdictions to meet commercial objectives (Khalil and Zayani, 2021); they invest more heavily in local content in some countries than others (Lobato and Lotz, 2021); and consumers face regional restrictions in content availability (Elkins, 2019). Long-standing concerns about the consolidation of power imbalances between countries and the risk of cultural imperialism have taken on renewed urgency (Lotz et al., 2022) in this context. Netflix continues to contend with, and claims to continue to try to optimise its recommender systems for, more localised country and regional preferences: including preferences related to language (Raimond and Basilico, 2016) – a key dimension of diversity.

In Australia, long-standing broadcast television legislation (Australian Parliament, 1999) dictates that 55% of programs broadcast on primary channels each year between 6 a.m. and 12 a.m. must be ‘Australian content’. When Netflix entered the Australian market in 2015, it was not constrained by this legislation, and was therefore able to attract large numbers of Australian subscribers through its premium offerings of US television (Lobato, 2019; Turner, 2018). Nevertheless, Australian viewers are accustomed to seeing high levels of Australian television (Deloitte, 2020), and to viewing diversity through an Australian lens (Khoo, 2022), creating demand for local content and complexity for recommender systems. For example, many Australians would expect that ‘Black’ programming will include representations of ‘Blak’ Australians – Aboriginal and Torres Strait Islander people (Turner, 2020).

It is important to note that identity-based diversity can operate in tension with nation-based customisation (or ‘localisation’). Global platforms are important sites for queer affiliation and belonging across geographic borders (Zhao, 2021); Indigenous creators and audiences may be deeply rooted in place while building global networks and resisting being positioned in relation to colonial nationhood (Fredericks et al., 2022).

Conclusion

The role of recommender systems in mediating diversity on behalf of global platforms is complex, fraught, and deeply consequential for content producers and creators, as well as for their publics: but the very fact that diversity has so many meanings and values attached to it make it a productive site for intervention. TikTok's public-facing recommendation guidelines¹ are instructive in this regard. At first glance, they suggest the company conceptualises ‘diversity’ in a narrow, easily operationalisable sense (recommending content by different creators or about different themes) because ‘[o]ur community tells us they love finding creators they wouldn’t have known to follow, or learning about a new interest because it was recommended to them, along with enjoying content that already matches their taste’. But the platform also makes a subtle appeal to rationales for diversity beyond user experience, stating that its ‘systems won’t recommend two videos in a row made by the same creator or with the same sound’, not only because ‘[d]oing so enriches the viewing experience’ but also because it ‘can help promote exposure to a range of ideas and perspectives on our platform’. The platform also positions recommendation ‘diversity’ as part of its approach to ensuring user safety: claiming to diversify the viewing experience of users away from content that ‘though not violative of our policies, could have a negative effect if that's the majority of what someone watches, such as content about loneliness or weight loss’. In that sense, ‘diversity’ is presented not just a way of engaging users, it is also an element of ‘responsible’ curation. Amid widespread concerns about recommender systems’ tendency to reinforce existing views and relationships rather than expose users to new ones (creating ‘filter bubbles’), and in the amplification of harmful ideas and modes of expression, TikTok promises to expose viewers to ‘a range of ideas and perspectives’ and shield them from content that may be ‘problematic if viewed in clusters’.

Diversity, then, is a heavily freighted keyword (Williams, 1977; see also Peters, 2016) in the global digital media environment: it is a multivalent concept that wields operative power in a variety of ways over the external policy environments, internal rules, content, and technical elements of platforms, with significant implications for the careers of cultural producers and experiences of audiences. Following Williams, we choose to see this play of meanings not only as a symptom of political struggle, but as a resource of hope, however modest, for productive interventions.

Recent advances in the field of recommender design, in combination with a range of pressures on platforms to improve the amount and quality of diversity in their recommendations, offer nascent possibilities. As the brief survey above suggests, the dimensions of diversity already used by recommender systems in cultural and entertainment platforms range across a matrix that mixes cultural, business and governance considerations: in many jurisdictions they need to balance the local, national, and global; in order to be effective in their own terms they need to ensure aesthetic and content diversity; and for both practical and ethical reasons, representative social and identity diversity may be part of the picture. To fully engage with these competing practical and normative questions is going to be challenging but increasingly necessary for global platforms. To do so would involve experimenting with multidimensional models of diversity in collaboration with diverse creator and user communities, and in dialogue with civil society.

Footnotes

Funding

The author(s) disclosed receipt of the following financial support for the research,authorship,and/or publication of this article:This article reports on research that was supported by the ARC Centre of Excellence for Automated Decision-Making and Society (ADM+S) and partially funded by the Australian government through the Australian Research Council (project number CE200100005);as well as funding from the European Research Council,under the European Union's Horizon 2020 Research and Innovation Programme,in the form of an Advanced Research Grant awarded to Professor David Hesmondhalgh at the University of Leeds (Grant agreement no. 1010020615).

ORCID iDs

Jean Burgess

Joanne Gray

D. Bondy Valdovinos Kaye

Ariadna Matamoros-Fernandez

Author biographies

Jean Burgess FAHA is Distinguished Professor of Digital Media in the School of Communication and Digital Media Research Centre,Queensland University of Technology,and Associate Director of the Australian Research Council (ARC) Centre of Excellence for Automated Decision-Making and Society. Her most recent book is Everyday Data Cultures (Polity Press,2022).

Louisa Bartolo is a PhD Candidate at the Digital Media Research Centre,Queensland University of Technology and a student member of the ARC Centre of Excellence for Automated Decision-Making and Society. She specialises in the governance of algorithmic recommender systems.

Joanne Gray is a Lecturer in Digital Cultures at the University of Sydney,a Chief Investigator at the International Digital Policy Observatory,and Commissioning Editor for the journal Policy & Internet. Her research seeks to understand how digital platform companies exercise private power,and to explore relevant policy options.

Jonathon Hutchinson is Chair of the Discipline of Media and Communications at the University of Sydney. He is President of the Australian and New Zealand Communications Association and Editor in Chief of the Policy & Internet Journal. His recent book,Digital Intermediation: Unseen infrastructure for cultural production,is available through Routledge.

D. Bondy Valdovinos Kaye is a Post-Doctoral Research Fellow at the University of Leeds researching digital music industries,music creators,and public policy. He completed a PhD in Creative Industries in 2021 at Queensland University of Technology,Australia.

Ariadna Matamoros-Fernandez is an ARC Discovery Early Career Research fellow (DE230101558),Senior Lecturer in the School of Communication at Queensland University of Technology,Chief Investigator at the QUT Digital Media Research Centre (DMRC) and Associate Investigator in the ARC Centre of Excellence for Automated Decision-Making and Society. Her research looks at social media cultures,online harm,and platform governance.

Kylie Pappalardo is an ARC Discovery Early Career Research fellow (DE210100525),a Senior Lecturer in Law and a Chief Investigator in the Digital Media Research Centre at the Queensland University of Technology,and an Associate Investigator with the ARC Centre of Excellence for Automated-Decision Making and Society.

Patrik Wikstrom is a Professor of Computational Communication in the School of Communication at QUT,Chief Investigator in QUT's Digital Media Research Centre and Associate Investigator in the ARC Centre of Excellence for Automated Decision-Making and Society. His research is focused on how digital technologies shape the consumption and production of digital culture.

References

Agrawal

Gans

Goldfarb

(2018) Prediction Machines: The Simple Economics of Artificial Intelligence. Cambridge, MA: Harvard Business Press.

Aguiar

Waldfogel

(2021) Platforms, power, and promotion: Evidence from Spotify playlists. Journal of Industrial Economics 69(3): 653–691.

Anderson

Maystre

Anderson

Mehotra

Lalmas

(2020) Algorithmic effects on the diversity of consumption on Spotify. In: Proceedings of The Web Conference 2020. New York: Association for Computing Machinery (WWW ’20), pp. 2155–2165.

Asmar

Raats

Van Audenhove

(2023) Streaming difference(s): Netflix and the branding of diversity. Critical Studies in Television 18(1): 24–40.

Australian Parliament (1999) Broadcasting Services Act 1992 (Cth), s. 121G(1) (legislation).

Bourreau

Moreau

Wikström

(2022) Does digitization lead to the homogenization of cultural content? Economic Inquiry 60(1): 427–453.

Cotter

(2021) ‘Shadowbanning is not a thing’: Black box gaslighting and the power to independently know and credibly critique algorithms. Information, Communication & Society 26(6): 1226–1243.

De Gemmis

Lops

Semeraro

and Musto C (2015) An investigation on the serendipity problem in recommender systems. Information Processing & Management 51(5): 695–717.

Deloitte (2020) Everybody gets it: The economic and social benefits of commercial television in Australia. Report for Free TV Australia, August. Available at: https://www.accc.gov.au/system/files/Free%20TV_1.pdf.

10.

Duffy

Meisner

(2023) Platform governance at the margins: Social media creators’ experiences with algorithmic (in)visibility. Media, Culture & Society 45(2): 285–304.

11.

Elkins

(2019) Algorithmic cosmopolitanism: On the global claims of digital entertainment platforms. Critical Studies in Media Communication 36(4): 376–389.

12.

Fredericks

Bradfield

Nguyen

and Ansell S (2022) Disrupting the colonial algorithm: Indigenous Australia and social media. Media International Australia 183(1): 158–178.

13.

Gerrard

Thornham

(2020) Content moderation: Social media’s sexist assemblages. New Media & Society 22(7): 1266–1286.

14.

Gillespie

(2022) Do not recommend? Reduction as a form of content moderation. Social Media + Society 8(3): 20563051221117552.

15.

Gomez-Uribe

(2016, February 17) A global approach to recommendations. Netflix. Available at: https://about.netflix.com/en/news/a-global-approach-to-recommendations (accessed 8 March 2023).

16.

Helberger

(2018) Challenging diversity – social media platforms and a new conception of media diversity. In: Moore

Tambini

(eds) Digital Dominance: The Power of Google, Amazon, Facebook, and Apple. Oxford: Oxford University Press, pp. 153–175.

17.

Helberger

Pierson

Poell

(2018) Governing online platforms: From contested to cooperative responsibility. The Information Society 34(1): 1–14.

18.

Hildén

(2022) The public service approach to recommender systems: Filtering to cultivate. Television & New Media 23(7): 777–796.

19.

Khalil

Zayani

(2021) De-territorialized digital capitalism and the predicament of the nation-state: Netflix in Arabia. Media, Culture & Society 43(2): 201–218.

20.

Khoo

(2022) Picturing diversity: Netflix’s inclusion strategy and the Netflix Recommender Algorithm (NRA). Television & New Media 24(3): 281–297.

21.

Kunaver

Požrl

(2017) Diversity in recommender systems – A survey. Knowledge-based Systems 123: 154–162.

22.

and Liu F (2020) Case-based reasoning for resolving the diversity-accuracy dilemma of recommendation system through subdivision. In IEEE, 2020 IEEE 3rd International Conference of Safe Production and Informatization (IICSPI), pp. 357–360.

23.

Lobato

(2019) Netflix Nations: The Geography of Digital Distribution. New York: New York University Press.

24.

Lobato

Lotz

(2021) Beyond streaming wars: Rethinking competition in video services. Media Industries 8(1): 89–108.

25.

Lopes

(1992) Innovation and diversity in the popular music industry, 1969 to 1990. American Sociological Review 57(1): 56–71.

26.

Lotz

(2021) In between the global and the local: Mapping the geographies of Netflix as a multinational service. International Journal of Cultural Studies 24(2): 195–215.

27.

Lotz

Eklund

Soroka

(2022) Netflix, library analysis, and globalization: Rethinking mass media flows. Journal of Communication 72(4): 511–521.

28.

Middlebrook

(2020) The grey area: Instagram, shadowbanning, and the erasure of marginalized communities (17 February). Available at SSRN: https://ssrn.com/abstract=3539721 or http://dx.doi.org/10.2139/ssrn.3539721

29.

Morales

Lamarche-Perrin

Fournier-S’Niehotta

, Poulain R, Tabourier L and Tarissan F (2021) Measuring diversity in heterogeneous information networks. Theoretical Computer Science 859: 80–115.

30.

Napoli

(1999) Deconstructing the diversity principle. Journal of Communication 49(4): 7–34.

31.

Nicholas

(2022) Shedding light on shadowbanning. Center for Democracy and Technology. Available at: https://cdt.org/insights/shedding-light-on-shadowbanning/ (accessed 8 March 2023).

32.

Noble

(2018) Algorithms of Oppression. New York: New York University Press.

33.

Peters

(ed.) (2016). Digital Keywords: A Vocabulary of Information Society and Culture, vol. 8. Princeton, NJ: Princeton University Press.

34.

Peterson

Berger

(1971) Entrepreneurship in organizations: Evidence from the popular music industry. Administrative Science Quarterly 16(1): 97–106.

35.

Raimond

Basilico

(2016, 17 February) Recommending for the world. Netflix Tech Blog. Available at: https://netflixtechblog.com/recommending-for-the-world-8da8cbcf051b (accessed 8 March).

36.

Schedl

Knees

McFee

, Bogdanov D and Kaminskas M (2015) Music recommender systems. In: Ricci

Rokach

Shapira

(eds) Recommender Systems Handbook. Boston, MA: Springer US, pp. 453–492. doi:https://doi.org/10.1007/978-1-4899-7637-6_13

37.

Spotify (n.d.) Spotify for developers – Web API. Available at: https://developer.spotify.com/documentation/web-api (accessed 8 March 2023).

38.

Steinberg

(2017) Introduction: Regional platforms. Asiascape: Digital Asia 4(3): 173–183.

39.

Turner

(2018) Netflix and the reconfiguration of the Australian television market. Media Industries 5(2): 129–142.

40.

Turner

(2020) Dealing with diversity: Australian television, homogeneity and indigeneity. Media International Australia 174(1): 20–28.

41.

Williams

(1977) Keywords: A vocabulary of culture and society. Science and Society 41(2): 221–224.

42.

Zhao

(2021) Introduction: Global queer fandoms of Asian media and celebrities. Feminist Media Studies 21(6): 1028–1032.

‘Diversity’ as multidisciplinary keyword for the politics of cultural recommender systems in global digital media platforms

Abstract

Keywords

Introduction

Diversity in recommender system design

Representational dimensions of diversity

Geographic dimensions of diversity

Conclusion

Footnotes

Funding

ORCID iDs

Author biographies

References