Streaming platform users choose their songs and tend to use search engines and artists’ discographies.
Based on a study involving a panel of 4,000 users selected at random on a music streaming platform over a five-month period, Jean Samuel Beuscart, Samuel Coavoux and Sisley Maillard, all researchers and sociologists at Orange, conducted a detailed analysis of the impact of algorithmic recommendation tools on their users.
Currently, around 80% of French online users listen to music via streaming. The success of specialised platforms like Spotify and Deezer, or more general platforms like YouTube are reflected in an increased share of income from streaming as part of music industry revenue. With 85% of digital income and 42% of overall market share, streaming is now the dominant way to consume music.
With several million songs available at our fingertips, these services offer a variety of algorithmic tools, ranging from a collaborative filtering tool (people who liked this song also liked that one), themed recommendations (other songs by the artist you listened to), to themed radio stations (featuring music inspired by this artist) or playlists “fully inspired by your tastes”, not forgetting the many “atmosphere”-based playlists tailored to users’ moods. Designed to help and guide listeners in their choices, the recommendation algorithms used in streaming services have a broad range which thus seems to meet their expectations.
To ensure success in this domain, Jean Samuel Beuscart, Samuel Coavoux and Sisley Maillard, all sociologists and researchers at Orange, conducted a study based on an analysis of the listening patterns of a panel of 4,000 users selected at random on a music-streaming platform and monitored this for a period of five months. It was the perfect opportunity to have a better understanding of the impact of algorithms on users’ music choices and to see whether or not the transition to digital listening had altered how they consumed, and above all, what they consumed.
It also provided a way to analyse whether or not these tools promote greater cultural diversity, as well as whether or not users actually follow the recommendations made by the algorithms, in which situations, how often and with what effects?
Principles of the study
The data collected provides information about user profiles (age, gender, place of residence, registration date, subscription type, favourite artists and songs), their consumption (date, duration, format used for listening) and characteristics of the content streamed (artist, song, release date). There is also the listening context variable, which helps to identify the type of device, recommendation or collection influencing a user’s decision to listen to a given song.
The platform on which the study was conducted offers a catalogue of 35 million songs and two unlimited offers via a “freemium” economic model: one is free and financed by advertising, the other attracts a fee of €9.99 per month. The fee-based package allows users to listen to their music in high quality, offline and without advertising. These offers are available on computers and mobile devices such as smartphones and tablets.
In terms of listener characteristics, it appears that the population tended to be young, male, and city-based. Only 37% of the sample were female. The median age was 28 and 80% of users were between 17 and 46 years old.
Most listeners were rather seasoned users: 53% of users joined the service more than two years ago and 18% joined more than five years ago.
Two thirds of the sample preferred to use their mobile phone to access the service, while one third used their computer. The type of subscription offer had an influence on the device used. In the case in point, subscribers to premium offers (with no advertising) used mobile phones more frequently and sometimes even exclusively, to listen to music, in contrast to free subscribers who tended to use their computer instead.
Variety, frequency and diversity of songs
To measure the impact of musical recommendation algorithms on listeners, this study analysed the intensity and diversity of streamed songs and how these were split between well-known and lesser-known artists.
Over five months, the sample’s online users streamed over 17 million songs. This figure must be weighted by the actual duration of streamed listening on the platform: just over half of the streams (56%) went right to the end of the song, while one third did not last more than 30 seconds.
These streams covered nearly one million different songs. Each time a song was streamed at least once for over 30 seconds (81%), it was listened to on average 16 times by users. This means that the average song was listened to once a month for every thousand active users. “However, this average is misleading because the distribution of the number of streams is unsurprisingly very asymmetrical: 39% of songs were listened to only once by a single user, while only 5% were streamed more than 50 times,” says Samuel Coavoux.
The most popular songs therefore garnered a considerable audience: the five most-streamed songs during the study totalled more than 25,000 streams each and each of them was streamed by almost half the listeners on the panel.
In terms of frequency of use, this is rather high since on average, users logged into the platform for 86 days during the study (i.e. one out of every two days) to listen to some 50 songs each day they logged in. These results again show high disparities, because the most active 10% of users listened to over 10,000 songs over the period, whereas the least active 10% listened to fewer than 450.
“The promise of these streaming platforms is that, given their extensive offering, with both free and flat-rate subscriptions, and their finely-tuned recommendation tools, music consumption should be more diversified than on the physical or download markets,” notes Samuel Coavoux. If we look at variety, i.e. the number of different artists streamed, we can see that the consumption of a user subscribing to a full offer with an almost unlimited range of songs results in greater variety than that observed for other ways to access music. “Even then, the comparison of this variety with other formats is tricky because the listening conditions are not the same and we cannot measure offline listening with the same degree of accuracy,” explains Samuel Coavoux. In fact, compared to the physical or download markets, streaming a song does not incur any additional expense and compared to radio, the consumer plays a more active role in choosing the songs they listen to and the offer is much less restricted. The study revealed that the number of songs streamed on the platform by just 4,000 users over five months represented approximately 9 times more songs than those broadcast by all French radio stations in the course of a year. However, a recent study by the French Ministry of Culture showed that this increase in variety benefited artists at either end of the popularity scale, to the detriment of those in between.
The variety of consumption also varied greatly from one individual to another. On average, a user listened to 1,149 different songs: 5% listened to three times that number, while one quarter of the sample listened to no more than 359. There was a correlation between variety and the age of the user. According to Samuel Coavoux, “This means, on the one hand, that younger consumers have less diverse listening patterns and on the other hand, that the most diversified consumption comes from more frequent users, who have more experience in using the platform, and are more inclined to use the recommendation tools – these users are older. They stand in contrast to younger users who listen to more well-known artists on their mobile phones.”
With 10% of artists accounting for 90% of streams, the study revealed that almost all individuals listen to songs by moderately popular artists. Niche artists represented only 17% of all artists listened to by the panel.
These differences are even more apparent if we consider streaming frequency rather than users’ music libraries: one in four users listened almost exclusively to highly-popular artists and it was more common to listen to few or no niche artists (53% of users spent less than 5% of their listening time on 90% of the least-streamed artists).
“In fact, popular artists dominate the music libraries of users whose music consumption is low, while also often representing the most frequent streams by more active users,” notes Samuel Coavoux.
Guidance and discovery
The authors of this study also analysed whether the various forms of guidance helped strengthen diversity of consumption, one of their promises, or whether it actually led to reinforced uniformity in terms of music listening?
The first observation was that, apart from “most listened to”-style charts, most recommendation systems tended to guide listeners to artists that were less well-known than the average first plays. First plays generally focused less on the most-streamed artists. However, this result varied greatly according to device: it was stronger with algorithmic recommendations, then with autonomous exploration of artists’ discographies and new releases. “Contrary to what you might think, autonomous exploration systems are not particularly cutting-edge, although they offer a wider range than editorial recommendations and charts,” notes Samuel Coavoux. However, these “discoveries” did not necessarily remain in the user’s collections: songs discovered via algorithmic recommendations received fewer replays than others during the observation period.
This study also helped to put into perspective the contexts in which listeners tend to stream music. “Some songs are streamed more often in a multi-activity context, when music may simply be playing in the background. We choose a playlist precisely so that we do not have to consciously decide to listen to it.”
The same was true of musical genres. The study showed that recommendations were unevenly distributed according to musical genre – some genres such as blues, jazz or dance seem to be better suited to guided listening. This is due to the fact that the use of recommendations depends on the listening contexts with which the musical genres are associated. We can see that some genres are particularly suitable for activities other than listening and are therefore more likely than others to be played in the background. For instance, dance music for parties, blues for a calm atmosphere, etc.
Guidance systems tend to be used more, in a non-exclusive way, by frequent users.
Above all, the results of this survey help to have a better understanding of emerging algorithmic cultures, in which listeners, according to their skills and situations, will reject algorithms, or use them on an ad-hoc basis without surrendering control of their musical experience. “Qualitative work will nonetheless be required to understand the contexts, the level of conscious listening, comprehension, type of attention, and the objective to which users agree to rely on algorithms,” concluded Samuel Coavoux.
This is one to watch…
In this study, several major types of recommendation tools were distinguished:
– personalised algorithmic recommendations rely on users’ previous listening history to calculate their preferences and suggest appropriate songs Among them, we can distinguish ad-hoc algorithmic recommendations relating to an artist (you liked X, you should like Y), an album or a playlist, from those composed via an algorithm that creates a flow of recommended songs until interrupted by the user: a continuous flow of personalised recommendations, an “artist radio station” that plays songs similar to those by the artist.
– editorial recommendations are produced by humans, experts or other users: themed playlists, editorial radios, etc.
The platform also provides tools such as charts (most listened-to artists, most streamed songs this week) and editorial selections (a playlist created by a given artist, DJ, or commercial partner).
Context-based browsing tools allow users to search for an artist via the search engine and browse an artist’s discography.
The last tool is based on explicit links built up between users who are “friends”