[Paper Review] The Impact of the Filter Bubble -- A Simulation Based Framework for Measuring Personalisation Macro Effects in Online Communities

Thomas Gottron, Felix Schwagereit|arXiv (Cornell University)|Dec 20, 2016

Opinion Dynamics and Social Influence11 references3 citations

TL;DR

This paper proposes an agent-based simulation framework to measure macro-level filter bubble effects in online communities due to personalization algorithms. Using topic modeling and network generation models, it compares content-based and author-based personalization, finding that content-based filtering reduces filter bubble severity by preserving broader social networks and vocabulary diversity, despite similar precision gains.

ABSTRACT

The term filter bubble has been coined to describe the situation of online users which---due to filtering algorithms---live in a personalised information universe biased towards their own interests.In this paper we use an agent-based simulation framework to measure the actual risk and impact of filter bubble effects occurring in online communities due to content or author based personalisation algorithms. Observing the strength of filter bubble effects allows for opposing the benefits to the risks of personalisation.In our simulation we observed, that filter bubble effects occur as soon as users indicate preferences towards certain topics.We also saw, that well connected users are affected much stronger than average or poorly connected users. Finally, our experimental setting indicated that the employed personalisation algorithm based on content features seems to bear a lower risk of filter bubble effects than one performing personalisation based on authors.

Motivation & Objective

• To measure the macro-level impact of personalization algorithms on information diversity in online communities.
• To investigate whether and how filter bubble effects emerge when users express topic preferences.
• To compare the risks and benefits of content-based versus author-based personalization on social network structure and vocabulary diversity.
• To identify which user groups are most affected by filter bubble effects in simulated online communities.

Proposed method

• The simulation uses an agent-based model with agents representing users who generate and consume messages based on individual topic interests.
• A generative network model creates social networks with varying degrees of user connectivity, including hubs.
• A state-of-the-art topic modeling approach assigns topics to messages, reflecting user interests and content relevance.
• Two personalization algorithms are implemented: one filtering based on message content features, and one based on the author's topic profile.
• The framework measures precision, recall, active social network size, and vocabulary diversity to quantify filter bubble effects.
• Simulation experiments vary parameters like core and peripheral interest probabilities to assess algorithmic impact under different user preference conditions.

Experimental results

Research questions

• RQ1 To what extent do personalization algorithms induce filter bubble effects at the macro level of online communities?
• RQ2 How do content-based and author-based personalization algorithms differ in their impact on social network structure and vocabulary diversity?
• RQ3 Which user groups—especially in terms of connectivity—are most vulnerable to filter bubble effects?
• RQ4 Can content-based personalization reduce filter bubble severity while maintaining high precision compared to author-based filtering?

Key findings

• Filter bubble effects emerge as soon as users indicate preferences for core topics, leading to a biased perception of information based on their interests.
• Well-connected users are significantly more affected by filter bubble effects than average or poorly connected users due to higher message exposure and algorithmic reinforcement.
• Author-based personalization leads to a stronger reduction in active social network size and vocabulary diversity compared to content-based personalization.
• Despite similar precision and recall values, content-based personalization results in less severe filter bubble effects, preserving broader social and linguistic exposure.
• The simulation shows that users can still access off-topic messages, suggesting that filter bubbles are not absolute and may be reversible through user behavior.
• The results indicate that content-based personalization may be a more balanced approach, minimizing social and cognitive isolation while maintaining user experience benefits.