When people ask me what my biggest vice is, I often respond: 'I watch too much YouTube'. And it's true. I spend enormous amounts of my free time on this incredible platform full of entertainment, randomness, and also education. Actually, when I graduated from high school, luckily with quite decent grades, my dad confessed to me that he had been very weary of my consumption of YouTube during my finals. I was aware of the issue but I always had good excuses for it all. First and foremost, I used YouTube extensively as a tool to learn about any random curiosity I had, watch TED talks and also to improve my English. But to be totally honest, back then, I definitely did not spend most of the time on this "good" side of YouTube. Way too often the YouTube algorithm took me down the rabbit hole, spitting me out after hours of videos of people making knifes out of chocolate or even more obscure materials. Various times I tried to reign in my use of YouTube, to focus it on only the educational part by removing subscriptions to channels with no educational value and sometimes even deleting the app from my phone altogether. However, my foe developed too and soon was able to draw on my urges for entertainment and distraction to coerce me to spend hours on hours on the platform starting with that one pesky video on the homepage, which was exactly clickbaity enough to get me to click it but not too clickbaity to alert me to the trap. And so our confrontation continued...
The revelation for the way out
In 2020, months into working for my current employer, I finally started to use a separate Chrome account for my work to get better separation of work-and-life. With this also came a fresh YouTube account. In the beginning, I then occasionally used this account to search and watch tutorials about coding or work related things, making sure to keep it exclusively to that. It was meant for work only, after all. Once I had amassed a bit of watch time on exclusively work-related things, I started to notice how the YouTube algorithm adapted to what I was watching. Suddenly it helped me find videos and lectures that taught me so much more than what I usually sought out myself. At the same time, during quarantining at home, I became increasingly aware of the needs I had throughout the day. Not the least thanks to a small habit tracking app I created. I noticed that often, after long stretches of meetings or deep work, my body was just longing for that bit of distraction or entertainment. But, even more importantly, I also noticed that not only the typical entertainment videos helped me fill this need but also the lectures, tutorials and other educational videos the "new" algorithm was recommending to me. Suddenly, a light-bulb in my head went off.
After years and years of fighting with the YouTube algorithm, maybe I could finally make it my ally by consciously training it to serve me the great educational videos and lectures I would never find myself?
Deconstructing the "phenomenon" I was seeing with my work account, coupled with the basic understanding of recommender systems and AI that I gained through the videos that this little bot recommended to me, I determined, that if I could only remove all of the rabbit holes and clickbaity videos I had previously fallen prey to, I could change the picture the algorithm had of my interests.
The logic behind the algorithm
A recommender system, a technical term for the algorithm driving the YouTube or Netflix recommendations we love to hate for being too good, is an algorithm employed to help users discover the content they are interested in. Essentially, it works like your good friend who you ask for a good movie to watch. Your friend will know from previous movies you watched together what movies you enjoyed and which not as much. He can then take two paths: Either he notices that this other friend of his seems to have a super similar taste as you do and that this friend just recently told him about that great new movie on Netflix he loved, thus he may recommend you to watch this. Alternatively, your friend may also try to look for the patterns in the movies you watched together before and notice that you really enjoyed long movies with this famous actor that play in the past. Thus he may think of another movie that matches this pattern and then recommend it to you. The more features of the movie he recommends adhere to your preferences, the surer he can be that you'll love this new movie, too.
In tech these system work by analysing the behaviour of the respective user on the platform. The way a user interacts with the individual pieces of content thereby helps the system to understand what they like and what not. As an example, if a user watches only 3 minutes of a 10 minute video, it signals to the system, that the user was not as interested into the video as a user who watched all 10 minutes of it. By combining this with various other data points, the system can then determine a score for the given piece relating to the level of interest a user has for the given video. Looking at this across multiple videos, computers will then be able to find patterns in the users watch patters. They then can apply similar techniques as your friend did, either by finding other people that have similar interests as you or by looking for the features that coincide with your interests. No matter the way the algorithm finally comes to his conclusion, however, the first requirement for the functioning of the algorithm is the abundance of personalised user data through interactions with the platform. If a user changes this or even deletes it, the algorithm is forced to start from the beginning.
How to "hack" your recommendation algorithm
If you want, as a user, you have the opportunity to take advantage of this fact. Thanks to the recent implementation of GDPR, online platforms like Google and Netflix have been required to give you the option to have your data deleted and you can use this to get the opportunity to press the reset button and have the system see you with completely fresh "eyes".
If you want to go ahead and start off with a clean slate, here's how you can do it:
- Make sure you have cleaned e.g. your subscriptions of undesirable channels. We will only delete our watch history, so the algorithm will only be able to go by the other pieces of information it has about you when it restarts its work. Thus it will take your subscriptions very serious when recommending new content to you, so you probably want to be on the save side.
- Navigate to your YouTube profile and the segment ("Your Data in YouTube") You should find something similar to this:
You want to click on "Manage your YouTube Watch History". You'll get to the page which shows you all of your Watch History.
-
Next up, you want to click on "Delete activity by".
-
Finally, you'll be presented with the Delete Activity Dialog. Here you want to click on "All time". Be sure though that this is your intention. There will be no way back and you'll need to retrain the algorithm again. It may take a bit.
When history was
You did it! Congratulations. You have left your history behind and can show the algorithm who your new you really is.
After this, you'll see that the recommendations you are receiving have probably really reduced in quality. This is only temporary, however, and improve with the amount of new signals you're sending to the algorithm. So make sure that you help your new buddy a bit by searching for the topics you want it to recommend to you, watching videos about it and marking videos it recommends but are not interesting to you as "Not interested" (you can do this by clicking on the three dots next to videos). You'll have to have some self-control but doing this now will help you to get tempted less in the future.
The Recommender Systems of the future
Recommender systems have long been a staple tool for social media companies to increase the amount of the new currency of our society - attention. They started using them without explicit user lobbying mostly optimising for their own profit but in the process they have built incredibly useful tools that can be used against the interest of users in a more mindful multimedia use but show similar promise for the use for good. The content oceans we nowadays find ourselves in have quickly grown too big for mere mortals to navigate and recommendation algorithms are the perfect technology to help us traverse this issue. However, left by their own devices the systems value each interaction of the user as an equivalently valuable insight. While this is perfectly in line with the idea of the user who just wants to stick to his own habits, many of us are also aware of the many instances in which we don't behave as we would like to. However, these mishaps get valued equally much as every conscious action we do. Similarly it gets increasingly difficult to convince the algorithm to support you in your new year's resolutions as you'll need to get the algorithm to reduce the value placed on your past self and increase it for your new one instead.
However, I believe magic can happen when we increase the interface with the algorithm to work in a similar manner as a shopping assistant in a shop, being able to tell it what we look for or of what we want more and disregarding what we don't. Maybe through something like this, we will finally build a collaboration with a machine that shows us the most beautiful corners of a world we cannot feasibly discover ourselves.