How are Artists Affected by Spotify’s Genre Classification?

A Comparative Study between French and British Playlists Produced by Spotify

Hanna AGBANRIN • Ziyu DENG • Yuzhe LIU • Toshali SENGUPTA • Fabian Bi Sheng SIAU
5 December, 2021

Introduction

On 5th June 2020, Republic Records, an entity within Universal Music Records, announced that they would ‘remove “urban” from the label’s verbiage in describing departments, employee titles and music genres’ (Aswad, 2020). This announcement came in the midst of a nation-wide conversation around systemic racism in the United States after the death of George Floyd, Ahmaud Arbery, and Breonna Tyler. #TheShowMustBePaused and Blackout Tuesday movements notably demanded that the music industry take accountability for its disparate treatment of Black artists, by ‘protecting and empowering the black communities that have made them disproportionately wealthy in ways that are measurable and transparent’ (idem). Republic Records was the first of many actors within the American music ecosystem to adopt such a measure and was notably followed by the National Academy for Recording Arts and Sciences which organises the Grammy Awards. The National Academy decided to undertake a similar change and substitute the category ‘Best Urban Contemporary Album’’ for ‘Best Progressive R&B Album’ (Lewis, 2020). This change was presented as a way to ‘appropriately categorize and describe this subgenre’ and ultimately use ‘a more accurate definition to describe the merit or characteristics of music compositions or performances themselves within the genre of R & B’’ (idem).

This movement gained traction in France, as the French streaming platform Deezer decided to follow suit. Deezer’s justification was built on a rationale similar to that of the Grammys – they saw the term urban as an outdated label that inadequately coalesces Black artists together, regardless of their music genres (Fanen, 2020). However, this decision was not as widely accepted in France as in the US. Many major stakeholders still use the term ‘urban’, including record labels and Spotify France. Notably, Spotify France has argued that the controversy around the ‘urban’ term is a mere ‘anglo-saxon debate’, which is not relevant to French society, as the ‘urban’ label adequately and accurately refers to the ‘music of the youth in urban areas’ (Cachin, 2020). This reflects the intrinsic difficulty to categorise and classify music. Musical genres are categories that are ever-evolving, reassessed and redefined, as artists innovate and experiment through genre-bending by mixing features associated with a specific genre, with features associated with another.

The debate in France is focused on the category ‘pop urbaine’ which is used to refer to hip hop, r&b, dancehall, afrotrap and ragga (Lemaître, 2020). Whilst some argue that pop urbaine is a relevant label – referring to a musical melting pot incorporating diasporic sonorities (from North Africa and Sub-Saharan Africa), the French local musical culture, and the US art form of hip hop (Cachin, 2020) – others see this term as an abstract and stigmatising label which underscores a disparate treatment between aesthetics associated with artists of colour and those of white artists (Fanen, 2020). In other words, whilst advocates see this term as a legitimate music category, critics view this genre as an outdated catch-all term which reflects a lack of understanding of the evolution of the musical scene in France amongst young artists of colour.

Our study is thus guided by the question How are artists affected by Spotify’s genre classification?, looking specifically at the case of France. It aims to contribute to this debate by investigating the implications of the use of the label ‘urban’ by Spotify France and in its music recommendation algorithm. Our findings show that the use of such broad catch-all categories leads to an unnecessarily heightened competition between artists, making it harder for niche and independent artists to be recommended by the algorithm to users. This sets the pop urbaine genre apart from more specific genres on Spotify France, but also from countries in which Spotify has attempted to develop a more specific, comprehensive and fine-grained understanding of the local music scene. Ultimately, this phenomenon results in the promotion of a select few well-established artists, and prevents Spotify from enabling the listener to discover new artists. Most significantly, specific and precise labelling is imperative for the emergence and discovery of new, independent, and nicher artists, and allows for a proper understanding and account of the evolution of music amongst young artists of colour.

Context

Spotify, currently the largest music streaming service in the world, is a Swedish start-up founded in 2016. The service bases its marketing strategy and its user experience heavily on the volume of user data and content data it collects. In our study, we focused on Spotify’s genre-categorisation algorithms and recommendation algorithms. These two systems collaborate by first identifying certain characteristics of a track, an album, or an artist, and then, based on those identified characteristics, recommending the music to the targeted audience. Here, we focus on analysing the possible discrimination towards artists brought about by these two Spotify algorithms. Similar to most profit-driven companies, Spotify has not made its algorithm transparent and openly accessible. However, it does disclose certain compositions of its algorithms from various channels. From here we have noted that Spotify adopts content-based and user-history-based algorithms.

Content-based algorithms categorise millions of tracks into scattered genres by examining features embedded within tracks and playlists. With content-based algorithms, Spotify has developed an Audio-Feature model with nine dimensions (Sree 2021), which are key, mode (major or minor), acousticness (a 0 to 1 measurement that indicates the confidence level in the acoustic nature of a track), danceability (a 0 to 1 measurement that indicates how suitable a track is for dancing), energy (a 0 to 1 measurement that indicates how fast, loud, and intense a track is), instrumentalness (a 0 to 1 measurement that indicates how instrumental a track is), loudness (measured in decibels), valence (a 0 to 1 measurement that indicates how positive a track sounds), tempo (measured in beats per minute), and popularity (a 0 to 100 measurement that indicates how popular a track is). Among these, there are dimensions that are originally measured numerically, while there are also dimensions that are transformed into numerical values. Thus, the fundamental rationale behind this Audio-Feature model is to systematically extract sensational and abstract features from music, and to convert them into quantified, standardised, and computational data.

Although a collection of articles dedicated to analysing the relationship between the popularity of a song and its Spotify Audio Features (Amsterdam 2019; Georgieva et al. 2018; Middlebrook & Sheik, 2019) already exists, very few studies have attempted to examine whether the Spotify machine learning empowered Audio Features can accurately classify music in a manner that conforms to general human perception, and also respects nuances among highly diverse genres. Apart from Audio Features, Spotify also employs natural language processing (NLP) as a part of its content-based categorisation mechanism. Spotify’s NLP algorithms constantly extract word-based information on the Internet, including songs’ lyrics, descriptions, labels, people’s posts and comments, and reviews from industrial publications, and thus develop more and more enhanced predictions of how people understand music and genres (Krysik, 2020).

Another component of Spotify’s algorithms on classification and recommendation system is based on user-history. People’s listening history, search history, playlists, etc, provide valuable information about their preferences, which enables Spotify to generate user-profiles of its audience and to thus recommend content tailored to their tastes (Pinarbasi, 2019; Melchiorre & Schedl, 2020). More importantly, Spotify applies collaborative filtering to its user population, which compares a single user’s profile to a large group of people, and produces clusters of people with similar preferences (Pérez-Marcos & Batista, 2018). Collaborative filtering studies existing behaviors and predicts future behaviors based on well-established group portraits. It thus improves its efficiency by moving from comparing similarities among music and individual users to comparing at a collective level (Johnson, 2014). Introducing user-history based algorithms is an important step for matching the right music for the right ear. Holding such a large volume of content data and user data, Spotify employs the mechanism called matrix factorisation to map the presumed correlations among users and content. Recognising the fact that each user has accessed a limited amount of content and only a very limited amount of users have accessed a large enough amount of content, matrix factorisation aligns factors concerning users and those concerning platform’s content in a weighted manner, in order to generalise the known alignments within the matrix to the overall population (Huq, 2019).

Successfully predicting the genre of music and therefore introducing it to the targeted audience is crucial for Spotify and also for artists, since both parties gain revenue when a track is played by a single user. Therefore, being categorised ‘properly’ means higher possibility of being discovered by the interested audience as well as the potential audience base, and if one’s works are selected into the representative playlist of a specific genre linked to Every Noise At Once – a Spotify-affiliated musical genre exploration website that maps thousands of existing popular and niche genres (Patch, 2016) – it definitely brings the advantage of visibility, and therefore, popularity. Although Spotify has already developed such a sophisticated categorisation and recommendation system, whether the system can produce unbiased accurate predictions or not is still up for debate. One of the most salient issues is inherent in its Audio Features, which calculates the popularity of a track as a presumably objective value. Since popularity is brought into the consideration of promoting a track or not in the first place, already popular tracks and artists may enjoy an even higher level of exposure, which sits at the core of our inquiry.

Literature Review

The proliferation of artificial intelligence (AI) has ‘led to a number of calls for “transparency”’ (Walmsley, 2021: 585) which is noted in multiple fields of AI such as translation algorithms, law enforcement, and more. Significantly, Floridi et al. (2018 in Walmsley, 2021: 587) describe the opacity of algorithms and their ‘machine bias’ in ‘socially significant’ contexts, which refers to a situation where individuals or groups are prejudiced by biases produced or exacerbated by AI, yet do not have any viable recourse due to the lack of transparency of the algorithm. Walmsley (2021) notes that on most occasions, machine bias problems are ‘identified and fixed with human intervention’ and thus it is critical that AI systems should be designed in a way that allows for ‘challenging or contesting automated decisions’ (588). Ultimately, this highlights a need for greater scrutiny of algorithms on which users ‘seem to preferentially trust and rely […] to make predictions and decisions in a growing range of socially significant and morally weighty contexts’ (idem: 594).

One of these socially significant contexts is that of the music industry, in particular streaming platforms like Spotify. Though music recommendations and classification might not appear to be as ‘morally weighty’ as that of the use of algorithms in judicial proceedings or crime prediction as discussed in the literature (Walmsley, 2021), it nonetheless has a profound impact on artists’ livelihood – especially independent artists and artists from minority groups (e.g. women) (Ferraro et al., 2021; Dominik et al., 2021; Shakespeare et al., 2020). Focussing on streaming platforms, Anderson et al. (2020: 1) found that ‘algorithmically-driven listening through recommendations is associated with reduced consumption diversity [and] that when users become more diverse in their listening over time, they do so by shifting away from algorithmic consumption and increasing their organic consumption’. This is unsurprising, considering the ‘curatorial power’ that digital platforms have in ‘inaugurat[ing] relations of dependency among creators and the industries they draw upon’ (Prey, 2020: 1). Platforms such as Facebook, YouTube, and Spotify ‘do not produce – or own – most of the content they circulate’ (ibid) but they have the power to decide what and how to present the content on their platforms through the ‘organizing and programming of content’ (idem: 8). A particular feature that streaming platforms leverage are playlists, which Prey (2020) argues that Spotify uses to ‘gradually reduce its reliance on record labels by slowly increasing the percentage of directly licensed or unsigned content within [playlists]’ (8). Wielding its curatorial power, Spotify is able to unilaterally (re-)organise music produced by artists with or without their input:

‘As a result, Spotify has become a “new gatekeeper” (Bonini & Gandini, 2019) and a “very powerful intermediary” (quoted in Shah, 2018)’ (Prey, 2020: 7).

However, Kamehkhosh et al. (2020) posit that ‘little is known so far about the effects of [playlist] recommendation functionality’, but found that the ‘mere presence of the recommendations impacts the choices of the participants, even in cases when none of the recommendations was actually chosen’ (285). They found that ‘track recommendations were highly adopted by [their] participants’ and ‘helped [them] discover relevant tracks and influenced their choices even when they did not select one of the recommendations’ (idem: 287). These findings suggest that playlists influence Spotify users, whether in an overt or subliminal way.

Within the realm of machine bias in streaming platforms, academics note that ‘gender fairness is one of [artists’] main concerns’ and the ostensible bias against independent labels, arguing that major labels are favoured (Ferraro et al., 2021: 249; see also Shakespeare et al., 2020; Dominik et al., 2021). In particular, Ferraro et al. (2021) argue that music recommendation systems ‘may privilege the content of a small group of artists when maximizing user satisfaction’, thereby ‘limits some artists’ chances to reach a larger audience particularly due to the feedback loop’ which refers to the algorithm learning to ‘recommend increasingly similar items’ (249). In the light of this potential bias, Shakespeare et al. (2020) cite ‘recent reports [of] the disproportionate treatment of female artists […] prevalent in the Western music industry’ (1) but found in their study of gender bias in Spotify playlists that there is a bias both for and against female artists but no ‘relations between [that] and the disproportionate low streaming share of female artists on the platform’ (2). Instead, they conclude that ‘recommender systems can propagate a pre-existing bias’ but may not necessarily ‘cause the emergence of new forms of biases’, albeit cautioning that their study made use of data that ‘come mostly from Western countries [which] cannot be generalized to represent a global scenario’ (idem: 8).

Dominik et al.’s (2021) study, on the other hand, found that ‘current recommender systems do not work well for consumers of beyond-mainstream music’ and are particularly ‘prone to popularity bias, which leads to the behavior that long-tail items (i.e., items with fewer user interactions) have little chance of being recommended’ (1-2). This implies that artists that produce music perceived to be non-mainstream would find it more difficult to be recommended to listeners compared to their mainstream counterparts. They conclude their study ‘validat[ing] related research by showing that beyond-mainstream music listeners receive a significantly lower recommendation accuracy than mainstream music listeners by four standard recommendation algorithms’ (idem: 22). This not only disadvantages non-mainstream listeners on Spotify but also the artists themselves who are unable to reach their target audience. On the surface, recommendation systems are supposed to help users ‘discover new products’, however, it might in practice ‘lead to a reduction in sales diversity’ and ‘prevent what may otherwise be better consumer-product matches’ (Fleder & Hosanagar, 2009: 697).

Given that ‘recommender systems have become an essential means to help users deal with content and choice overload as they assist users in searching, sorting, and filtering these extensive music collections’ (Dominik et al., 2021: 1) and that ‘playlisting algorithms tend to privilege certain genres over others’, the workings and effects of music recommendation systems ‘represent a serious social justice concern’ (Chodos, 2019: 39) that requires closer scrutiny.

Methdology Flowchart

Playlists considered: Pop Urbaine, Who We Be, GrandHit

Since the starting point of our study was the broad genre of Pop Urbaine, we considered the Spotify created Pop Urbaine playlist as our primary dataset, which claims to include “all pop urbaine and afropop hits”. To allow a comparison between two similar playlists, we included Who We Be, the UK’s biggest “Hip-Hop, Afrobeats, Dancehall and RnB playlist”. Since we wanted to explore the intersection between hip hop and French music, the comparison with Who We Be allowed us to study this across language and geographical region. However, these two playlists did not account for variances that might arise from English music in Spotify being better categorised. This is why we also included GrandHit, a playlist that contains the 50 most popular French hits of the moment. By using GrandHit as a comparison with Pop Urbaine, we were able to clarify whether the issue lay with categorisation of French music or specifically Pop Urbaine.

Methodology

We have represented our methodology in the flowchart presented. To collect the data we either used the Spotipy API to generate dataframes of playlists or used the website OrganizeYourMusic to generate tables that we formatted into excel sheets to later import into our notebook. Most of our analysis was done using Python and the Pandas library, and some of the data visualisation was done using Data Wrapper. We decided to analyse three types of datasets from these playlists. First, we looked at the frequency of artists within playlists. Second, we looked at the frequency of genres and subgenres within the playlists. Finally, we considered the frequency of artists and their popularity from the radios of random artists associated with the three main playlists. These steps are better explained in the notebook, especially with regards to the code used and the specific dataset referred to. The logic underlying our analysis was counting and creating visualisations to show frequency and variance.

Analysis

All of our analysis is conducted in the following notebook: Spotify Data Analysis
The link gives you the ability to comment on the notebook. The datasheets required to run the code have also been uploaded to this respository.

Discussion and Conclusion

Our study has demonstrated that there is a difference in the exposure of popular artists in Pop Urbaine compared to Who We Be and Grand Hit. Aguiar et al. (2021: 1) underscore the significance of playlists, arguing that they are the ‘main mechanism for promotion’ on streaming platforms like Spotify and Apple Music. Having control over the curation of playlists provides Spotify the ‘editorial capacity to transform the industr[y] they intermediate’, with Spotify ‘heavily promoting the playlist format since at least 2012’ (Prey et al., 2020: 1). In this spirit, Spotify has been ‘using its editorial capacity to promote its own playlists over playlists created by the major labels and other third parties’ (ibid), supported by the fact that some major record labels ‘have ownership stakes in Spotify, which could give Spotify a reason to provide more advantageous promotion of their products’ (Aguiar et al., 2021: 1) – creating unfair competition systemically rigged against smaller and independent labels. This leaves Spotify users without opportunities to diversify their listening, restricting their consumption to a commercial mainstream, through which niche artists do not stand a chance of increasing their exposure without being featured on an editorial (Spotify) playlist.

With regard to the logic behind editorial playlists, Chodos (2019) examined ‘hubs’ on Spotify. Hubs refer to collections of playlists that are loosely tied together by a common ‘theme’, which might be a musical ‘genre’. Chodos (2019: 51) predicates his argument on the ‘traditional idea of genre […] that there are certain musical properties shared among all members of a genre’, referencing Spotify ‘hubs’ to problematize the notion of genres on the platform. In an attempt to illustrate the arbitrary nature of Spotify ‘hubs’, he cites how they are ‘significant for their extra-musical references’ such as ‘hubs’ for ‘study’, ‘sleep’, and the American celebrity ‘Ellen’ alongside ‘hubs’ such as ‘reggae’ which might be considered a musical genre (ibid). This is congruent with our findings which demonstrate an inconsistency in genre labelling as well as seemingly haphazard labels such as ‘francoton’, which refers to most tracks in two playlists that are presented as being different by Spotify. This term only exists on Spotify, and is apparently a portmanteau made of France + Reggeaton. It would seem that the francoton label is a catch-all term which may be a geographical indicator, signaling French songs, or a genre which regroups genres like variété française, indie pop, pop urbaine, rap. Our findings also demonstrate that even though in France there seems to be less genre differentiation, there is a significant difference between a monolithic ‘pop urbaine’ almost entirely constituted of francoton, and a more varied ‘variété française’ where Francoton is the dominant genre, but other genres co-exist such as chanson and french indie pop.

To improve current systems, some academics suggest the development of newer models of music recommendation and categorisation systems. For example, Zhang (2021) – in response to traditional music categorisation algorithms – explores music categorisation using ‘closely related sound spectrum features’ which are ‘combined with one-dimensional convolution, gating mechanism, residual connection, and attention mechanism […] based on convolutional neural network[s]’ to remedy the problems associated with ‘manual and traditional machine learning methods’ (8). Nonetheless, Sturm (2013) cautions that ‘neither classification accuracy, nor recall and precision, nor confusion tables, necessarily reflect the capacity of a system to recognize genre in musical signals’ (371), arguing that ‘the recognition of genre is to a large extent a musical problem, and must be evaluated as such’ (373). In other words, algorithms in isolation – no matter how accurate they may be – are inadequate in genre classification, as they generally ‘do not reliably produce genre labels indistinguishable from those humans produce’ (idem: 400). This could potentially suggest the need for industry consultations to ensure a fair and just representation of artists’ musical genres, especially for underrepresented artists. Aguiar et al.’s (2021) study of the manually curated Spotify playlist ‘Friday’s New Music’ found that ‘despite concerns about bias against independent-labels and women, [their] results indicate that independent music, and music by women, receive better ranks than their eventual on-platform streaming performance seems to warrant’ (3), leading them to suggest that ‘Spotify may be responsive to the criticism […] lead[ing] [Spotify] to actively promote work from the groups voicing concerns’ (22), whilst acknowledging that this specific playlist is not representative of Spotify as a platform. Nevertheless, the supposed responsiveness of Spotify towards industry and public criticism might allow for consultations between the platform and artists.

Furthermore, our study contributes towards (the limited) research comparing music industries between different countries, as we had compared playlists from Spotify France and Spotify UK. Bauer & Schedl (2019: 29) indicate that it is important to distinguish between ‘mainstreaminess on a global and a country level’, proposing ‘three groups of countries: (i) those countries where users’ music consumption behavior corresponds to the global mainstream (e.g., the United Kingdom (UK), the United States (US), the Netherlands (NL)), (ii) those countries that show a distinct country-specific mainstream that is listened to in addition to the global mainstream (e.g., Finland (FI), Brazil (BR), Russia (RU)), and (iii) those countries where the global mainstream is important in the country but, still, users listen very frequently to some artists that are not part of the global mainstream (e.g., Japan (JP), China (CN), Indonesia (IN))’. In our study, France would likely fall in group (ii) and the UK in group (i). Future research should take these differences into account, perhaps by measuring ‘country-specific differences in music listening behavior with respect to the degree of deviation from the global mainstream’ (ibid).

Whether or not one can produce an algorithm that is accurate, neutral, and truly satisfies all stakeholders remains a contentious question and a constant work in progress, as Chodos (2019: 52) advises, ‘recommendation algorithms are far too intimately personalized, too frequently updated, and too complex’, and that one has to ‘bear in mind that [one’s] conclusions are almost always based on incomplete and possibly outdated information’.

References

Aguiar, L., Waldfogel, J., & Waldfogel, Sre. B. (2021). ‘Playlisting favorites: Measuring platform bias in the music industry’ International Journal of Industrial Organization, 78(C). Available at https://www-nber-org.acces-distant.sciencespo.fr/papers/w29017 [Accessed 14 November 2021].

Amsterdam, Noah. (2019). ‘ Analyzing popular music using Spotify’s Machine Learning Audio Features.’ Digital Humanities @ Kenyon. Available at: https://digital.kenyon.edu/dh_iphs_ai/14/

Anderson, A., Maystre, L., Mehrotra, R., Anderson, I., & Lalmas, M. (2020). ‘Algorithmic Effects on the Diversity of Consumption on Spotify’ WWW’20, Taipei, Taiwan, 20-24 April.

Aswad, J. (2020). Republic records to cease using ‘Urban’ term. Variety. https://variety.com/2020/music/news/republic-records-urban-term-black-music-1234626400/ [Accessed 3 December 2021].

Bauer, C. & Schedl, M. (2019). ‘Global and country-specific mainstreaminess measures: Definitions, analysis, and usage for improving personalized music recommendation systems’ PLoS ONE, 14(6): e0217389.

Cachin, O. (2020). La pop urbaine existe-t-elle ?. RFI Musique. https://musique.rfi.fr/rap/20201218-pop-urbaine-existe-elle [Accessed 3 December 2021].

Chodos, A. T. (2019). ‘What Does Music Mean to Spotify? An Essay on Musical Significance in the Era of Digital Curation’ INSAM Journal of Contemporary Music, Art and Technology, 2(1): 36-64.

Ferraro, A., Serra, X., & Bauer, C. (2021). ‘Break the Loop: Gender Imbalance in Music Recommenders’ CHIIR’21: Proceedings of the 2021 Conference on Human Information Interaction and Retrieval, Canberra, Australia, 14-19 March: 249-254.

Fanen, S. (2020). À la benne, Les « musiques urbaines ». Les Jours. https://lesjours.fr/obsessions/musique-racisme/ep5-musiques-urbaines/ [Accessed 3 December 2021].

Fleder, D. & Hosanagar, K. (2009). ‘Blockbuster Culture’s Next Rise of Fall: The Impact of Recommender Systems on Sales Diversity’ Management Science, 55(5): 697-712.

Georgieva, E., Suta, M., & Burton, N. (2018). ‘Hitpredict: Predicting hit songs using spotify data.’ Stanford Computer Science 229: Machine Learning. Available at: http://cs229.stanford.edu/proj2018/report/16.pdf

Huq, P. (2019). ‘Music to My Ears: De-Blackboxing Spotify’s Recommendation Engine.’ CCTP-607 Big Ideas in Tech: AI to the Cloud. Available at: https://blogs.commons.georgetown.edu/cctp-607-spring2019/2019/05/06/music-to-my-ears-de-blackboxing-spotifys-recommendation-algorithm/

Johnson, C. C. (2014). ‘Logistic matrix factorization for implicit feedback data.’ Advances in Neural Information Processing Systems, 27(78): 1-9.

Kamehkhosh, I., Bonnin, G., & Jannach, D. (2020). ‘Effects of recommendations on the playlist creation behavior of users’ User Modeling and User-Adapted Interaction, 30: 285-322.

Kowald, D., Muellner, P., Zangerle, E., Bauer, C., Schedl, M., & Lex, E. (2021). ‘Support the underground: characteristics of beyond-mainstream music listeners’ EPJ Data Science, 10(14). Available at: https://doi.org/10.1140/epjds/s13688-021-00268-9 [Accessed 14 November 2021].

Krysik, A. (2020). ‘What is the Secret of Spotify Recommendation System Success?’ Recostream. Available at: https://recostream.com/blog/what-is-the-secret-of-spotify-recommendation-system-success [Accessed at 16 November 2021].

Lemaître, F. (2020). L’histoire du terme ‘‘musique urbaine’’ et pourquoi il doit mourir vite. Brain Magazine. https://www.brain-magazine.fr/article/brainorama/61829-Lhistoire-du-terme-musique-urbaine-et-pourquoi-il-doit-mourir-vite [Accessed 3 December 2021].

Lewis, S. (2020). Grammy awards renames controversial “urban” category. CBS News. https://www.cbsnews.com/news/grammy-awards-rename-controversial-urban-category/ [Accessed 3 December 2021].

Masters of Media (2021) “Do fans also like…”? — in what ways do big artists profit from Spotify’s recommendation system. Available at https://mastersofmedia.hum.uva.nl/blog/2021/10/29/do-fans-also-like-in-what-ways-do-big-artists-profit-from-spotifys-recommendation-system/

Melchiorre, A. B., & Schedl, M. (2020). ‘Personality correlates of music audio preferences for modelling music listeners.’ Proceedings of the 28th ACM Conference on User Modeling, Adaptation and Personalization: 313-317.

Middlebrook, K., & Sheik, K. (2019). ‘Song hit prediction: Predicting billboard hits using spotify data.’ arXiv preprint arXiv:1908.08609. Available at: https://arxiv.org/abs/1908.08609

Patch, N. (2016) ‘Meet the man classifying every genre of music on Spotify — all 1,387 of them.’ Toronto Star. Available at: https://www.thestar.com/entertainment/2016/01/14/meet-the-man-classifying-every-genre-of-music-on-spotify-all-1387-of-them.html [Accessed at 16 November 2021].

Pérez-Marcos, J., & Batista, V. L. (2017). ‘Recommender system based on collaborative filtering for Spotify’s users.’ In International Conference on Practical Applications of Agents and Multi-Agent Systems: 214-220.

Pinarbasi, F. (2019). ‘Demystifying Musical Preferences At Turkish Music Market Through Audio Features Of Spotify Charts.’ Turkish Journal of Marketing, 4(3): 264-279.

Prey, R. (2020). ‘Locating Power in Platformization: Music Streaming Playlists and Curatorial Power’ Social Media + Society, April-June: 1-11.

Prey, R., Esteve Del Valle, M., & Zwerwer, L. (2020). ‘Platform pop: disentangling Spotify’s intermediary role in the music industry’ Information, Communication & Society, DOI: 10.1080/1369118X.2020.1761859

Shakespeare, D., Porcaro, L., Gómez, E., & Castillo, C. (2020). ‘Exploring Artist Gender Bias in Music Recommendation’ Proceedings of the ImpactRS Workshop at ACM RecSys ’20, Brazil, 25 September.

Sree, Sunku Sowmya. (2021). ‘Spotify–Song Prediction and Recommendation System’ Medium. Available at: https://medium.com/swlh/spotify-song-prediction-and-recommendation-system-b3bbc71398ad [Accessed 16 November 2021].

Sturm, B. L. (2013). ‘Classification accuracy is not enough: On the evaluation of music genre recognition systems’ Journal of Intelligent Information Systems, 41: 371-406.

Walmsley, J. (2021). ‘Artificial intelligence and the value of transparency’ AI & Society, 36: 585-595.

Zhang, J. (2021). ‘Music Feature Extraction and Classification Algorithm Based on Deep Learning’ Scientific Programming, 2: 1651560. Available at: https://doi.org/10.1155/2021/1651560 [Accessed 14 November 2021].