Algorithms are less in tune with hard rock or hip hop than with commercial pop
One of the most applauded characteristics of the platforms of streaming music like Spotify, Last.fm o Deezer is your topic recommendation system. It helps you navigate your extensive catalogs and, if it works well, it even allows you to discover new favorite bands. A recent study published in the magazine EPJ Data Science concludes that the type of music that is listened to influences, and quite a lot, in the effectiveness of the recommendations. More specifically, the work establishes that those interested in “conventional music” (the most popular or listened to) tend to receive recommendations that are much more suited to their tastes than those who prefer more alternative genres, such as hard rock or hip hop.
This problem has to do with the so-called popularity bias: songs with fewer interactions are less likely to be recommended by the users themselves, and therefore to be taken into account by the algorithm. It is also known from previous research that those who listen to music that escapes from the lane tend to have more complex user profiles: they accumulate more different artists listened to than lovers of popular music.
Therefore, the paradox arises that the users who generate the most data on musical interactions for the portals are the ones that get the least performance from the recommendation engine. Why?
The recommenders work is apparently simple: the algorithms record and classify all the music that each user listens to (musical genres and subgenres, name of the bands, average playing time, etc.). When someone clicks on a topic, the system shows them what like-minded users have heard.
The team of researchers who signed the study, all from Austrian and Dutch universities and research centers, decided to test the effectiveness of the recommendation systems to find the seams. To do this, he applied a computerized model to his own database made from the history of music listened to by 4,148 users of Last.fm. Half of them were chosen to be regular listeners of mainstream music, or mainstream music (the most common), and the other half prefer more alternative genres (those that deviate from the most common).
Four alternative subgenres
The more than 3.4 million songs included in the database were classified based on a series of “acoustic components” that describe the content of a certain track, the same ones that Spotify uses in its song screening systems: yes it is danceable, if the song includes voices or is entirely instrumental, if it is a studio song or played live, and so on.
With that information, the researchers obtained a more detailed portrait of what mainstream music is and what is not. From there, they used a computerized model to classify music that escapes the canons (the unconventional one) into four main categories: folk (music with a lot of acoustics), hard rock (music with a lot of energy), ambient (music with a strong acoustic and instrumental component) and electronics (a lot of energy and strong instrumental component).
After having all the users classified into subgroups, four different algorithms for music recommendation were applied. The study concludes that listeners who prefer background music receive significantly better recommendations than lovers of the hardrock.
The variety and quantity influence
The so-called “openness” (if lovers of a musical genre also recommend songs from other genres) and “diversity” (how different is the music that users listen to) are two variables that, the authors assure, have already been proven that influence the quality of the recommendations. One of the questions that the study set out to answer was whether listeners of unconventional music also have more open and diverse tastes than the rest, which would help to explain the lack of success of the recommendation algorithms.
When relating the results to openness and diversity, ambient music listeners turn out to be the most open but least diverse group (they recommend more songs from other genres, but listen to fewer bands), while with hardrockers the opposite happens: they are the least open to recommending other genres, but they follow more bands.
“Four recommendation algorithms have been applied and, in the four cases, the group hardrock it is the one that obtains the worst results. Those of the ambient music group, on the other hand, received better recommendations even than those of conventional music ”, underline the researchers, who call on the platforms to make an effort to better fit the tastes of those who listen to hard rock. “We understand that improving recommendations for this active group of users has other effects, such as improving artist exposure to recommendation systems.”