Stripe Partners
Spotify
Spotify
[/s2If] [s2If current_user_can(access_s2member_level1)] [/s2If]
This contribution is a case study of Spotify, a popular music streaming app, which uses automated recommendations to provide a better user experience to its listeners. Automated recommender systems have mostly been built around understanding user needs and user goals. Our case study presents a meaning-oriented approach aimed at understanding what users regard as meaningful and how an automated recommender system can forge meaning and offer experiences that help develop existing connections to music and generate new ones.
Following the meaning-oriented approach inspired by Lucien Karpik (2010), we were able to better understand how different audience segments engage with music and experience music as meaningful. We identified 2 cultural engagement models that listeners use to relate to music: (1) musical engagement during which music is the focus of the experience; and (2) non-musical engagement, during which the listener is the focus of the experience. Each engagement model uses different types of cognitive and evaluative aids, which we refer to as cues and proof points, to derive meaning from listening experiences. We also identified nine distinct types of experiences of meaning defined by distinct types of cues and proof points.
The proposed approach is applicable to the study and innovation of experience-led digital platforms and recommender systems.
Keywords: meaning, recommender systems, music, streaming
Article citation: 2020 EPIC Proceedings pp 191–202, ISSN 1559-8918, https://www.epicpeople.org/epic
Please sign in OR create a free account to access our library—the leading collection of peer-reviewed work on ethnographic practice. To access video, Become an EPIC Member.
Scale is a particularly urgent theme when researching and designing for digital platforms, algorithmic technologies and the attention economy. Some streaming platforms, such as Spotify, Netflix and YouTube, are based on business models which require them to acquire millions of users and provide value by creating customized, engaging experiences. In order to do that, these interfaces need to be automated so they can harness data to offer personalized content relevant to users’ tastes, contexts and moods.
In order to provide delightful listening experiences in every session for every listener, Spotify faces specific challenges and opportunities related to the affordances of its main medium, sound. Spotify is one of the unique applications where most of the user experience happens through people’s ears, brains and bodies, as opposed to their eyes. This poses several challenges when trying to understand what the value of user experience is and how to improve it.
First, unlike visual interfaces, where in-app interactions like time on the screen, likes and saves are the behaviors that could be used as proxies for understanding the value that users are deriving from the product, a lot of listening sessions on Spotify are hands and eyes free. This means that once users hit ‘Play’ and listen, we know very little about what experience they’re having. On top of that, any behaviours with the visual interface are highly driven by the context that listeners are in. For example, activities like running and driving a car, by their very nature, prevent people from interacting with the visual interface.
Second, based on the attitudinal segments, different types of users have different expertise in music and different abilities to navigate music, find music they like and discover new music. Their metaphors, expectations and benchmarks for deriving value from listening to music differ significantly and therefore, they require different forms and levels of support and feedback.
Thirdly, music listening itself is contextual. A playlist that is relevant at work might have a totally different meaning when listened to with kids at home and the measure of value evolves continuously with context.
Finally, while Spotify had a good understanding of what users’ needs are in various contexts, scenarios and use cases, there was a gap in understanding what they experience as meaningful. Our challenge was to understand how and why users derive meaning from music and how we might train the recommendation algorithms to respect that nuance of human experience.
Therefore, the underlying research challenge was: how do we scale automated recommender systems to forge meaning and offer content that helps develop existing connections to music and generate new ones?
FRAMING THE PROBLEM
In thinking beyond user needs and Jobs To Be Done (see, for example, Ulwick 2016), we were inspired by the sociologist Lucien Karpik and his book Valuing the Unique: The economics of singularities (2010). In this work, Karpik argues that cultural products such as music, wine, novels and movies, are singularities – complex, multidimensional goods, the value of which can’t be reduced to their specific features. It would be silly to claim one song has more value because it is longer, or because the singer hits higher notes. Or that a glass of red wine should be more expensive because it is a darker hue. Focusing on features in isolation misses the point.
Because value cannot be easily assigned to singularities, markets of singularities rely on complex mechanisms that enable actors to make decisions and choices and navigate uncertainty. Whereas in markets of commensurable goods, actors compare costs and benefits, in markets of singularities, they rely on what Karpik calls judgement devices and trust devices. Judgement devices “act as guideposts for individual and collective action” (Karpik 2010, 44) by providing cognitive support and opinion. Examples include reviews, charts or personal recommendations (Karpik 2010, 44–54). Trust devices help remove, dissipate or suspend uncertainty (Karpik 2010, p. 56) because they are often part of larger symbolic systems, such as social norms or formal authority. In the case of singularities such as movies or wine, we rely on the movie critic or wine connoisseur (and their training, education or expertise) to tell us what to expect, guide us in refining our tastes, teach us how to articulate the nuanced differences in our experience and ultimately, they help us make judgements about what is good and what isn’t, what we like and what we don’t. Because of the cultural complexity of singularities, we rely on these devices to serve as proxies of value.
Music is a type of singularity. It’s a type of product that requires knowledge and tastes for us to be able to make a judgement and choice about what music to listen to. Historically, radio has played an important role for the segment of listeners who are not confident in their ability to make choices about music by offering listening experiences curated by radio hosts who navigated the uncertain cultural field for audiences while also providing non-musical engagement, entertainment and information.
To continue differentiating itself as the leading audio streaming platform, Spotify needed to find ways to become a better ‘judgement device’ and a ‘trust device’ for songs, to use Karpik’s terminology, and to be able to do this in an automated way. However, to do this, we needed to move away from thinking about user needs and to understand what a meaningful experience of music can be.
Needs Versus Meaning
The true meaning of singularities only emerges to the user when they experience them themselves. Unlike user needs or goals, the value of singularities can’t be fully anticipated in advance. For example, Spotify’s previous research using the Jobs To Be Done framework, had identified that most listeners listen to music for fulfilling one or more Jobs (e.g. helping them focus, helping them change their mood, helping them create an ambience etc.). However, the Jobs To Be Done framework does not help in understanding how the value emerges for the listener as the listening experience progresses. For example, two users may derive entirely different meaning from the same good. One person could connect to a song because it soundtracked a breakup while someone else could love the same song because it gets people dancing. Jobs To Be Done framework and need-oriented approaches in general, miss this very important nuance.
Karpik explains that this uncertainty about what is valuable, which is a result of the incommensurability of cultural goods, is the defining characteristic of the market of singularities and requires an entirely different approach to value. In markets of commensurable goods, a consumer, Homo economicus, can make choices based on their needs and the expected costs and benefits. Their satisfaction is then derived in terms of efficiency. In markets of singularities, “Homo singularis must juggle the discovery, interpretation and evaluation of judgement devices; the discovery, interpretation and evaluation of singularities; sometimes the discovery, interpretation and evaluation of his own tastes; and a reasonable use of scarce resources.” (Karpik 2010, 67)
Following Karpik’s distinction between a need-oriented approach and a meaning-oriented approach has enabled us to come up with an entirely different model for thinking about the role that Spotify needs to play for its users and how this should be scaled throughout the organisation. We suggest that a meaning-oriented approach is more suitable for application to any services that offer cultural goods, such as music, video, film, fashion and luxury products, because it opens up opportunities to provide not only personalized experiences but also more relevant and meaningful experiences.
To illustrate the difference between the two approaches, consider a listener who may want to listen to the song Run the World by Beyoncé to improve their mood and feel motivated. Following a need-oriented approach, the job to be done is to enable them to search for the song, find it and play it as quickly as possible and without unnecessary friction to avoid frustration. Following a meaning-oriented approach might reveal that the listener experiences the song as meaningful because they identify with the archetype of a strong woman that this song represents. The recommender system could then songs by other artists who represent the same archetype, such as P!nk.
This distinction between a need-oriented approach and a meaning-oriented approach has implications for research design and analysis. While a need-oriented approach benefits from focussing on jobs to be done, the meaning-oriented approach benefits from the exploration of various judgement and trust devices that people use in order to navigate singularities. These can include recommendations from friends, popularity of artists but also personal aspirations. memories, travel experiences or social norms. A quote from Ken, 32, illustrates the complexity of music experience. Ken cannot easily identify his need or a goal. The meaning of the experience unfolds as his music is enjoyed by others, he gets complimented on it and helps him make new friends.
‘Honestly it’s whatever sounds good to me, that I can imagine myself at the beach listening to it I put it on, or anything that sounds good that I think other people will like… It creates a nice vibe, in the beginning I like it because it hypes you up to volleyball. A lot of people have said my playlist has been good, or they like my playlist and this is one woman in the group keeps on saying, I love your music because she use to be a DJ and then she is like oh! Can you share your playlist… I try to incorporate music that everybody likes so I have everything in there, and it’s constantly updated… It’s good for making friends, makes everyone happy.’
The proposed meaning-oriented approach suggests how we might help users like Ken make choices so they can have a more meaningful experience on Spotify.
Table 1. Comparison of needs-oriented analysis and meaning-oriented approach
| Needs | Meaning |
|---|---|
| Pre-existent – they drive choice | Emergent – can’t be anticipated in advance |
| Binary – are either met or not | Multiple – experienced in multiple, unpredictable forms |
| Choice is rational and based on calculation of expected costs and benefits | Choice and actions are justified when meaning is present |
| Meeting needs does not affect identity | Meaningful experiences affect individual’s identity |
| Systems are judged on efficiency | Systems are judged on the quality and relevance of the meaning |
The table is inspired by Karpik
METHODOLOGY
To tackle this challenge, we conducted ethnographic research with 12 participants in Boston in 2019, representing 2 types of audience segments:
- Lean in users are knowledgeable about music, understand their own tastes, have the vocabulary to articulate their preferences and are confident in discovering new music. These users understand musical genres and remember artists and songs. An example of how a lean in user can express his musical preferences:
Listening for the beat, I’m listening for the lyrics, what the artist is actually saying – do they flow on the beat, does it sound good together. Yeah, pretty much a good beat and then like good lyrics can win me over if executed well. (Noah, 25)
- Lean back users are not confident in understanding established categories, such as genres, nor are they confident in their own tastes. They struggle to remember or articulate what they like and rely on others to help them discover new music. An example of how a lean back user expressed her attitude towards music:
I’m more a radio person. So when it comes on the radio I’ll listen to it but I would say music is not something I’m like super obsessed … Like I enjoy music, I like it. But some people are always listening to music and always want to search for their own music, create their own playlist like they have a particular taste whereas I’m very much like fine with usually what’s on the radio. (Jennifer, 30)