March 7, 2019
We are flooded with choices. Everyday we have to make decisions about what food we buy, what movies we watch, and what music we listen to on our commute. Capitalism is firing on all cylinders and producing more and more options. All this can lead to serious decision fatigue .
Luckily, capitalism also provided its own solution to the problem: recommendation algorithms. Or also known as the technology that prioritizes what products, music, videos etc. is pushed in front of our eyes, ears, and, ultimately, our wallets. These are systems that help us sift through the vast array of possibilities at our disposal. They can be found in Amazon, Netflix, Youtube, Spotify, and many others. They are our guides in the digital content realm. Without them, it would take us much more time and effort to make a decision. Or at least, that’s what we are lead to believe. We don’t know for sure whether these systems actually recommend us things we need or want. It’s possible they simply lead us to consume whatever is presented to us.
Bear with me for the technicalities. These are essential to understanding how recommendation algorithms affect our decisions and the repercussions of relying on such systems.
The definition of recommender systems, which is the technical term for recommendation algorithms, is “information filtering system that seeks to predict the preferences of a user” . This can be achieved in two basic ways:
First of all, there are two basic types of recommender systems: collaborative filtering and content-based filtering.
Basically, collaborative systems rely on the network of users to make recommendations. In it, users usually rank products (in Spotify, ranking is usually achieved by the number of streams of a particular song). If two users have similar rankings of the same products, that means they have similar tastes. Thus, the algorithm can recommend products from the first user to the second, knowing that there is a high probability that the second will enjoy it.
You can think of it as a “bottom-up” approach, where the users dictate the behaviour of the algorithm. The advantage of this approach is that there is no need for supervision and that tastes occur organically. However, the disadvantage is obvious when a new product is introduced into the system. At the beginning it won’t have any ratings from any users, so it will be difficult to recommend. In the field this is referred to as the “cold start” problem .
The other possibility, content-based filtering, relies on experts. Thus it’s a “top-down” approach. These experts usually devise a system of categories for the products (in music, these would be genres). Then each of the products in the system are catalogued accordingly. When a user shows interest in a given category (by “liking” several musicians in a genre) the algorithm will suggest other products in that group.
The disadvantage here, from a technological point of view, is that producing such a system is very expensive and time-consuming. Pandora has invested significantly into this approach, and produced the Music Genome Project , a vast catalogue of 450 music genres and attributes.
While these are indeed the basic variants of the recommender systems, most companies nowadays use hybrid proprietary approaches that are not usually disclosed to the public. It’s the “special sauce” of the industry.
Now that we’ve established how they work, I would like to spend time on what they actually mean. How have these algorithms shaped our music behaviours?
Discovering new bands on Spotify, Google Music, Pandora, etc. with ease is one side of the coin. However, there are several consequences that come with the way we are presented new music, we should all be aware of.
Firstly: are these really new bands we’re listening to? Recent research reveals interesting consequences of recommender systems . On a personal level, we do indeed encounter new artists when going down the “rabbit hole” of recommendations. However, on an aggregate level, these algorithms seem to lead us all to the same artists. This in turn creates an effect of the “few rich get richer”. This is because popular items are usually presented first in the list of recommendations. Few people jump to the end of the line to choose the underdog .
Secondly, all machine learning and data analysis (which is what recommender systems rely on) resort to categorization of the data. In this case the data is us, the users. It’s only in this way that they know how and what to sell to us. Thus, all the users are reduced to proper little boxes. Think of Facebook’s advertising categories : they pigeonhole people into labels that are only there in order to optimize advertising. This happens in all tasks related to data analysis. The vast variety of human types will baffle any system, just like the variety of products baffles the human mind. In a sense, we get what we ask for. We want easy categorization of our products, then we ourselves become categorized by the system in a simplified way.
This is not by itself a bad thing: the human mind resorts to simplifications on a daily basis. It’s just that we need to be mindful to not get trapped in our own bubbles.
There are no easy conclusions to be drawn here. But I would like to remind you that we sometimes need to pop that bubble. One way I personally do this is by randomly choosing a genre on Last.fm  and then hopping from it to other related genres and artists that I’m curious about. In this way I use the system of tags and genres to discover new sounds.
Another way to do this is by exploring the local music scene around you. There are so many upcoming and underground acts that are amazing but will hardly ever make it to you through an algorithm. Coincidentally! Low-Fi has just that - plenty of intimate concerts that were created by local musicians and hosts. AND this is a great way to meet fellow music lovers and be part of a community that is built on respect and appreciation for music and the people who make it.