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.
Nowadays, there is increasing discussion about algorithms when talking about music and music video streaming services. Terms like “video metrics”, “clicks”, “average viewing times”, “comments”, “shares”, etc. are used. It is rare to talk about what the algorithms in these services are, and why they have reached their current position, which some estimate to be overly dominant. In some countries, streaming services and their algorithmic functions have also been the subject of investigations by competition authorities, particularly concerning their “mystique” and “opacity”.
Next, I will go through what the most common algorithm-driven functions in the music industry streaming services are, how they work, and what one should know and understand about them. I will not delve into the technical characteristics of the algorithms or their “technical or commercial fine-tuning”, but only general functions.
The article does not delve into the other curation of streaming services, where mainly individuals make choices about which songs end up on novelty or recommendation lists. The activities performed by the algorithms of streaming services are often referred to as algo-curation. Depending on the service, the weight of algorithmic curation and other curation can vary significantly. Understanding the algorithms helps to better comprehend the changes that have occurred in the commercial distribution of recordings and the business logics of record sales over the past few decades. Today, streaming services account for about 65 – 70% of the overall revenue stream in the recorded music industry. After this article, it is hoped that there will be a greater understanding of why particularly the metadata of recorded music and its various attachments are important. The article does not cover the various payout models currently in use by the services, which also affect music rights holders in many ways.
What are algorithms in general? – a brief and technical definition: an algorithm is a detailed, software-coded description or instruction that outlines how a task, process, or problem-solving is to be carried out. Algorithms are not tools or devices but rather descriptions of rules and processes encoded in code, as well as various chains of commands. In practice, we talk about learning algorithms and machine learning. The core characteristic of algorithms in streaming services is that they are self-learning, meaning the rules used by the algorithms change and evolve constantly based on the data they process. Additionally, algorithms are often ‘designed’ to also adopt broader ‘external’ trends. So, if a new genre or style becomes popular, the algorithms will automatically begin to offer related content to those users who are mathematically calculated to be a new potential target audience for various reasons.
Algorithms can be briefly defined as a precisely defined set of rules according to which customized content is selected and offered to the service user in various situations. And because the processes are technical, they are fast, and the advent of artificial intelligence functionalities has made the processes even quicker. Self-learning primarily relies on the data that an individual end-user listens to or watches, or correspondingly on what an individual service user explicitly does not want to listen to or watch.
The algorithmic functionalities used in services can be classified in many ways. One internationally used basic classification is as follows:
The data that the algorithms utilize is fundamentally quite simple in classification, namely:
The best impact is achieved when the song’s metadata is complete at the time when the song is delivered for distribution to the streaming service. This particularly involves various classifications, genres, styles, and performance languages of the song during the initial stage.
Secondly, it is important to create a network using the song and its metadata, meaning which is best and most profitable to offer the song to in the service. No song or video recommendation for individual users is coincidental, even though a suggestion of a ‘new and quirky’ song or video might sometimes seem that way to an individual user of the service.
The direct searches performed by the user provide exactly what is desired, but services also remember these and eventually link them to various functionalities that algorithms carry out. Thus, every search performed in the service ultimately connects to recommendations or at least indirectly affects recommendations or their absence for either that specific user or a friend.
An essential part of the current landscape is that a very large portion of listening and viewing is based on recommendations generated by algorithms used in the services, rather than users making precise song searches based on their own preferences. Conversely, this is something that is increasingly being considered in the professional marketing and distribution of recorded music. It is also public knowledge that algorithms operate differently in services depending on whether it is a free service with ads or a paid subscription model.
Generally, it is believed that the most advanced algorithms for music services can be found today on YouTube. The reason is clear. YouTube began actively developing its own algorithm system, or rather its precursors, nearly 15 years ago. Similarly, Netflix started using its first functions that could be classified as algorithms back in 2007.
In 2012, YouTube made watch time used for viewing videos an important metric. Before this, the most important factor was merely the number of videos stars. In 2015, YouTube implemented actual algorithms, largely as they are defined today. This also included more precise personalization and profiling of users of the service. Nowadays, practically every streaming service employ algorithms in its operations, each in its own chosen manner. For the end user, these functionalities appear to be various automatic and increasingly sophisticated recommendations and search results.
Nowadays, the algorithmic functions of other music services, such as TikTok or Spotify, are quite close to each other in terms of the “big picture”, and the deeper differences can be very minor. However, the comparability of the services is becoming more difficult, as the functionalities of algorithms in different services are increasingly kept within the realm of trade and business secrets, making it very hard to achieve transparency and general comparability. In the future, the situation will become more complex as the functionality of the services’ algorithms is based more and more on artificial intelligence functions.
Next, a short overview of how YouTube’s algorithms generally work. YouTube is indeed an exception among many services because it has a relatively transparent algorithm policy. Additionally, it has been studied extensively. However, Deezer and Spotify have also started to reveal more about how their algorithms work. Whether this is due to increased activity from competition authorities or a general and broader demand for transparency is difficult to say. Services justify their opacity not only with trade secrets but also by arguing that if rights holders and music creators were more familiar with how the algorithms of the services operate, it would lead to music production being influenced specifically by how to get music onto the service’s recommendation lists.
On YouTube, algorithms initially generate candidates for the user from the service’s content. The content formed this way is also ‘ranked’ in order of merit based on various values, weights, and variables. So basically, first a few hundred candidates are sought from millions of different content items, and then about a final dozen are listed from these hundreds. In practice, this process step works in such a way that the preferences of the end user that can be inferred from various factors are first determined.
In the next stage, different reasoning and predictions are reduced, and we move more into the realm of facts, meaning that search, listening, and viewing history, as well as the user’s possible likes, help to create the final ranking list for the content selected for recommendation. At this point, YouTube has already mathematically calculated various probability, and it is certain that the videos that end up on the ‘final list’ are based on the available data likely to be pleasing and ‘logical’ for the end user.
After this, there comes what is colloquially referred to as a “YouTube special” or “personal mix” on YouTube – this encourages the end user to also watch slightly different content – including somewhat less popular songs or videos. The aim is to find so-called “triggers” for many new recommendations. More about these triggers later.
YouTube’s algorithms are characterized by activity. The algorithms can be tuned to “scan” and actively search in various directions all the time. In summary, this works in such a way that if there is a so-called “wall” mathematically speaking, a new direction is taken. That is, if a certain genre does not work sufficiently for a user, the service actively and automatically starts searching for and finding different content and offering new suggestions.
