Have you ever had the strange feeling that your phone was listening to you? Have you ever felt that it knows you so well that it keeps showing you exactly what you like, again and again, until you barely see anything else? Then you have already experienced the power of recommendation algorithms. Their goal is to enhance your experience by predicting which content is most relevant to you. But sometimes that personalization can go a little too far. Your recommendations start looking more and more alike, the same ideas keep coming back, and other perspectives seem to slowly disappear from view. This gradual narrowing of the content and viewpoints you are exposed to is called a filter bubble.
If you have ever felt caught in one, chances are you blamed the algorithm — and you were partially right. But what if the algorithm were only half of the story? What if the choices you make, what you click on, watch or ignore, also help shape the recommendations you receive next?
This thesis explores exactly that interaction between users and recommendation algorithms. Rather than looking at filter bubbles as something created by algorithms alone, it studies how they emerge from the continuous feedback loop between what a system recommends and what a user chooses. The research follows three main objectives: observe, measure and mitigate filter bubbles. First, it investigates how recommendation algorithms evolve in response to different user behaviors, including through controlled experiments on YouTube. Then, it develops new ways to measure when recommendations are becoming increasingly narrow. Finally, it explores how recommendation algorithms could adapt their level of diversity to help users discover a broader range of content without sacrificing the relevance of their recommendations.
Ultimately, this leads to one final question: if our behavior can contribute to creating a filter bubble, could it also become part of the solution?