In a previous post, we tried to drill down to the true meaning of “serendipity”. We mentioned that serendipity is often seen as a means of “breaking the filter bubble”. But this raises yet another important question: What is a filter bubble?
Similar to serendipity, it seems 'filter bubbles' are difficult to pin down. That’s why in recent work, our researchers Lien Michiels and Jens Leysen from the Adrem Data Lab at the University of Antwerp have tried to come up with a measurable definition of the filter bubble.
But first, let’s go back in time to Eli Pariser’s filter bubble hypothesis once more. In 2011, Eli Pariser wrote a book on an idea that had been on people’s minds for quite some time already. Namely, that if we give more ‘power to the machines’ and allow algorithms to curate what we see online, won’t they only show us more of what we already know and love, and won’t that eliminate all potential for serendipity and discovery? And in doing so, wouldn’t we feed the further polarization and compartmentalization of our societies? And if all of that were to happen, wouldn’t this be a huge threat to democracy?
His book, “The Filter Bubble: What The Internet Is Hiding From You” caused quite the commotion, as did his TED talk on the subject. Suddenly, the filter bubble was at the top of everyone’s minds (and remained there for some time to come).
Pariser’s book and TED talk also sparked a flurry of research on the topic. Many studies have attempted to measure ‘filter bubble effects’ on social media platforms, news aggregators and other online platforms, looking at things like viewpoint diversity, diversity of topics or the spread of harmful misinformation. Others have tried to simulate what might happen on these platforms if we’re not careful.
Overall, the evidence is mixed. Simulation studies and sockpuppeting audits of online platforms, where users are impersonated by bots to exhibit specific behaviors, have shown that under the exact wrong conditions, users may be exposed to a very narrow information diet, that may consist of harmful misinformation. Studies with real users have come to milder conclusions: Most users seem to be exposed to sufficiently diverse content.
What we found most interesting when scouring the web for studies on filter bubbles, is that no one really seems to agree on what a filter bubble is. In our paper, we give an overview of empirical and conceptual work of the filter bubble and discuss the many different interpretations the concept has had.
Drawing inspiration from all these different works, we arrive at a ‘unified definition of the filter bubble'.
So what is a filter bubble then?
To start, we follow Dahlgren’s lead and separate the technological claims of the filter bubble from the societal claims. Whereas Pariser claims that the way recommender systems and search engines filter content will inevitably lead to increased polarization and compartmentalization of our societies, many have since contradicted this claim. Even if recommender systems and search engines would only show us more of what we already know and believe, people have lives beyond their browsers. They go to work, they watch TV, read a newspaper, practice sports, music, and other hobbies … All of which have the potential to expose them to a rich information diet.
These claims of polarization and compartmentalization are what we name the societal filter bubble. The behavior of recommender systems and search engines on the other hand is what we name the technological filter bubble. It is this technological filter bubble that we’ve tried to define.
We say that a technological filter bubble is "a decrease in the diversity of a user's recommendations over time, in any dimension of diversity, resulting from the choices made by different recommendation stakeholders."
In other words: a user finds themself in a technological filter bubble when over time, they are exposed to less and less diverse information.
Importantly, this can happen for many reasons. The recommendations and search results shown to a user are the result of a complex interplay between different stakeholders involved in the process. The user shapes their own recommendations by their behavior on the site and the recommender system is limited by the available content.
What does this mean for our understanding of filter bubbles?
First, it means that any analysis of technological filter bubbles requires nuance and we should steer clear of simplistic explanations that put the blame for creating filter bubbles squarely with the algorithm. Second, as users of these systems, we need to be aware of our own biases and the choices we make in selecting and consuming content.
As part of the Serendipity Engine project, we aim to gain a better, more nuanced understanding of how decisions made by different recommendation stakeholders might impact the ‘bubbliness’ of recommendation algorithms. Concretely, this means we will develop methods to measure the diversity of what is shown to users by different recommendation algorithms (under different conditions) and develop algorithms that lead to more serendipity and discovery on online platforms.
Does this spark your interest?
We hope to meet you along our way, serendipitously or meticulously planned. You might increase the chances of staying informed about the Serendipity Engine project by subscribing to our newsletter or connecting on LinkedIn.
Any questions in the meantime? Feel free to contact us at serendipityengine@vub.be.
This post is written by Adrem Data Lab (University of Antwerp). Adrem is mainly responsible for the work package on Recommender Systems. Adrem's main researchers on this project are Jens Leysen and Lien Michiels.
This text is intended for a general audience; if you are looking for a more in-depth discussion of some of the concepts, please refer to the linked academic publications. You may also want to check out the publications of our research team.