Search results
(1 - 1 of 1)
- Title
- Understanding and Combating Filter Bubbles in News Recommender Systems
- Creator
- Liu, Ping
- Date
- 2022
- Description
-
Algorithmic personalization of news and social media content aims to improve user experience. However, there is evidence that this filtering...
Show moreAlgorithmic personalization of news and social media content aims to improve user experience. However, there is evidence that this filtering can have the unintended side effect of creating homogeneous ``filter bubbles'' in which users are over-exposed to ideas that conform with their pre-existing perceptions and beliefs. In this thesis, I investigate this phenomenon in political news recommendation algorithms, which have important implications for civil discourse.I first collect and curate a collection of over 900K news articles from over 40 sources. The dataset was annotated in the topic and partisan leaning dimensions by conducting an initial pilot study and later via Amazon Mturk. This dataset is studied and used consistently throughout this thesis. In the first part of the thesis, I conduct simulation studies to investigate how different algorithmic strategies affect filter bubble formation. Drawing on Pew studies of political typologies, we identify heterogeneous effects based on the user's pre-existing preferences. For example, I find that i) users with more extreme preferences are shown less diverse content but have higher click-through rates than users with less extreme preferences, ii) content-based and collaborative-filtering recommenders result in markedly different filter bubbles, and iii) when users have divergent views on different topics, recommenders tend to have a homogenization effect.Secondly, I conduct a content analysis of the news to understand language usage among and across various topics and political stances. I examine words and phrases used by the liberal media and by the conservative media on each topic. I first study what differentiates the liberal media from the conservative media on each topic. I then study common phrases that are used by the liberals and the conservatives on different topics. For example, I examine which phrases are shared by the liberal articles on guns and conservative articles on abortion. Finally, I compare and visualize these words using different clustering algorithms and supervised classification methods.In the last chapter, I conduct an extensive user study to find possible solutions to combat the filter bubbles in the political news recommender systems. I designed a self-contained website that enables a content-based news recommender system and indexed 40,000 U.S.~political articles. I recruited over 800 U.S.~participants from Amazon Mechanical Turk (approved by IRB). The qualified participants are split into control and treatment groups. The users in the treatment group are provided transparency and interaction mechanisms, which grant them more control over the recommendations. Our results show that providing interaction and transparency a) increases click-through rates, b) has the potential to reduce the filter bubbles, and c) raises more awareness about filter bubbles.
Show less