Algorithms are increasingly driving what music we listen to on streaming services. In this article, Tuomas Talonpoika, Director of Gramex, looks at how algorithms work, what kind of data they use and why understanding them is important for every music professional.
Very important functionalities of the algorithms are the so-called linking and networking functions. In practice, this means linkiAn extremely significant functionality of algorithms is the so-called linking and networking functions. In practice, this means linking together various factors that connect different end users in different ways.
Next, we will focus on the two most used linking factors and metrics in streaming services, namely content and end user behavior.
1) Similarity of content: basic pattern – you receive suggestions for hip hop artists, for example, because you have also previously listened to some hip hop. The various information and identifiers in the metadata of the songs are important in this functionality.
2) End-User behavior: A good example of behavior is the ‘similarity of artists’ that have been listened to and viewed previously: as a devoted Elton John fan, you are recommended artists like Dua Lipa, because most other users who frequently listen to Elton John also often listen to Dua Lipa. In the streaming world, this is referred to as similar behavior. Services therefore actively monitor the connections between different artists and bands among other end users.
Most streaming services utilize a combination of content and behavior. Technically, this is relatively easy to implement. However, this combination involves the so-called ‘secret recipe’ of practically every service. That is, how and to what extent different combinations are made, and with what emphasis, is kept as a service-specific trade secret.
The so-called content similarity is usually the most important factor when new music is released, and no conclusions can be drawn at that stage regarding the popularity of the new material in relation to users’ final interest. This situation often correlates with the fact that the service lacks any usage or user analytics related to the content. This is often referred to as the so-called “cold start problem“.
In these situations, the importance of basic metadata is particularly highlighted; for example, genres, various sub-genres, and performance languages are important factors because they can act as a trigger throughout the entire service and its different users. There is also the downside that if a video “takes off” poorly, its recommendations may remain weak.
So, a spark must somehow be ignited. In these situations, streaming services usually add their own various types of “recommended metadata” to the track in addition to the metadata provided by rights holders or DSP aggregators. The services typically have full rights to do this according to their own terms of agreement. Generally, this addition happens based on the service’s own user analytics and profiling. Often, for example, just the lyrics, the language of the performance, or genre information can have a significant impact on how this so-called “cold start” begins.
The more challenging task is then the behavior and connecting it to recommendations. In these situations, the service’s analytics and algorithms must be able to find something interesting from previously occurring events in the service. In these situations, the algorithms sift through extensive usage data to determine what the broadest theoretical basis is among those end-users who enjoy certain types of songs or videos. The simple pattern is that the service begins to offer the user recommendation and playlist related to each user’s personality profile, largely based on the user’s own history primarily through songs or artists. In practice, systems go much further – in this case, the service’s algorithms attempt to combine the profiles of other users in various ways
So how can an individual music rights holder benefit from this? This raises the question of whether there is something that connects users among different artists. As mentioned above, do Elton John fans also listen to Dua Lipa? The same analytics can also be performed through cross-analysis of the genre classifications of different songs. And the more varied the similarities found in the service’s content regarding both songs and users, the better chance a song has of getting onto various recommendation lists through algorithms.
So, what works simplest from the perspective of the music rights holder? – one must be included in those playlists and recommendation lists that are listened to by users who enjoy and appreciate your music and style. This also involves the ‘hook’ that these other users are likely to listen to other similar artists, and thus algorithms gradually learn to understand which ‘group’ your music belongs to. This, in turn, helps you get included in recommendations and playlists that are increasingly closely related to your style.
The ‘network building’ of this listener market progresses step by step through the algorithms used by the service as follows:
▪ actual fans
▪ other users in the same market
▪ other users in similar markets
▪ other users in completely different markets
Various recommendation lists also include different risks, which I won’t go into detail about now, but it’s worth noting as an example that if a song gets included in new music recommendations and playlists, it can be a great and nice thing in itself, but in the world of algorithms, the situation is that various new release lists typically feature songs representing very different styles and genres. These may not necessarily have anything in common with your music. This situation is referred to in the streaming service world as ‘too early exposure’.
In practice, this means that a new song never reaches the users and recommendations based on content or behavior who would otherwise be the best and potential audience based on your style. It is likely that the algorithms primarily recommend your song to users who listen to completely different music on the same new release chart, for example. There shouldn’t be anything wrong with this, but in the world of algorithms, it likely leads to you being recommended to listeners of completely different genre categories. This carries the risk that you are unlikely to be wanted in that “group”, meaning you will be quickly and often dismissed. In the world of algorithms, this then decreases your popularity, and at the same time, it also decreases the so-called “long-term potential” of your other content. Therefore, it is worth remembering that the services and the algorithms within them remember very well everything that happens in the service – ‘both good and bad’.
One may naturally wonder how music streaming services really operate today with their recommendations. At this point, it should be noted that there are globally perhaps around 700 to 800 million paying subscribers across various music streaming services today. In addition, there are significantly more so-called free users. The share of paying subscribers is estimated to be about 1/3 of all music service users, depending on different sources of information. In these scales, getting on a recommendation list with even a 1% share creates quite a significant potential ignition surface for the fact that one of these approximately 20 million recipients of recommendations may decide to listen to or watch your songs. This, in turn, generates new impetus through the various ways users interact with each other and those who ‘network’ with them.
In 2020, it was estimated publicly that around 70% of the music videos viewed and listened to on YouTube are based entirely on content recommendations created by algorithms rather than so-called direct searches by users. Currently, the proportion of recommendations on YouTube is already much larger. It is believed that these figures may be somewhat lower in other similar music streaming services, but they are still high in any case. Additionally, music streaming services now include more and more various AI applications aimed at the users of the service, including creating different playlists. In these operations, algorithms play a significant role, and the latest AI-based functions work and create different playlists even with purely text prompts. Naturally, the services will analyze these prompts in the same way as any other usage and user history of the service.
From the perspective of music rights holders, the rise of recommendations, algorithms, and AI functionalities to a so-called key position is a very significant and important factor to consider when planning content licensing, for example, to YouTube as part of the compensation and revenue streams for music and recordings.
Additionally, various services increasingly include different cross-linkages, as is the case with iTunes and Apple Music. In other words, if a song is purchased heavily on iTunes, it may also appear in the growing recommendations of Apple Music.
One major change is the emergence of new players in the field, namely various so-called “DSP service promotion agents” – there is not yet a formal Finnish professional title for them. This group already constitutes quite a large segment. Likewise, new world distributors, meaning skilled aggregation services.
Similarly, the importance of distributors in the new world, that is, skilled aggregation services, has become even more pronounced – those who know and understand the rules and patterns of this ‘new world’ at least in outline are able to play the commercial game better than ever. We have therefore moved quite a distance from the traditional music sales revenue model, where a music fan buys a physical recording of their favorite artist or band from a music store.