About the project.
The goal of the project is to understand how people consume news on social media - and help them find content that represents multiple points of view about the issues they care about.
Why it matters
We believe that public dialogue and plurality of opinions are necessary for democracies to deal with societal issues -- and the World Wide Web has the potential to facilitate both. Nevertheless, much of the content we consume on the Web today is brought to us by algorithms, which might or might not be tailored to increase our awareness of the different viewpoints around important issues. If not, there is a real possibility that each of us consumes content that reinforces a specific point of view - perhaps one that we've already subscribed to. Such situations could lead to polarization and estrangement between societal groups that endorse different points of view but have little understanding about the arguments of each other. In this project, our goal is to generate content recommendations that alleviate such situations.
How does it work?
We build algorithms to automatically curate and recommend content to users who are involved in a political discussion on Twitter.
We first identify users that were active on the day after the latest presidential election (Nov 9, 2016). We classify these users into Democrat-leaning or Republican-leaning based on their retweet activity, using methods from our previous research. These methods essentially use the assumption that Democrats and Republicans predominantly retweet other Democrats and Republicans, respectively.
Subsequently, we identify content that is predominantly associated with either side - posted by prominent partisan accounts like @hillaryclinton and @berniesanders for Democrats or @realdonaltrump and @foxnews for Republicans; or posted mainly by accounts from one side.
Finally, we generate article recommendations for each user according to the following criteria:
- how well the topic of the article (represented with a set of TagMe entities) matches the interests of the user, based on past user activity;
- how likely we think the user is to endorse the viewpoint reflected in an article, based on the sides associated with the user and the article.
Eventually, for each user who participates in the study, we generate two recommended articles that are both close to their interests -- one closer to their viewpoint than the other.
What's next?
We will analyze how users from different sides evaluate different recommended articles. Details of our methodology will be made available as a pre-print and submitted for peer-review. Aggregate findings will be made available on a public webpage.
Data Policy
All data collected in this project will be used solely for aggregate analyis. No individual information is shared or disseminated. All data will be destroyed after the completion of the project. All recommendations are machine generated. We do not hand curate links for any specific account. The project is Open source and our code is available on request.
References
- Garimella, K., De Francisci Morales, G., Gionis, A. and Mathioudakis, M., 2016, February. Quantifying controversy in social media. In Proceedings of the Ninth ACM International Conference on Web Search and Data Mining (pp. 33-42).
- Garimella, K., De Francisci Morales, G., Gionis, A. and Mathioudakis, M., 2017, February. Balancing Opposing Views to Reduce Controversy. In Proceedings of the Tenth ACM International Conference on Web Search and Data Mining.
Contributors
The project is part of a continuing research effort on polarization on social media and consists of the following members.
- from Aalto University, Finland: Kiran Garimella, Michael Mathioudakis, Aristides Gionis.
- from Qatar Computing Research Institute, Doha, Qatar: Gianmarco De Francisci Morales.
For questions/comments, please contact Kiran Garimella (kiran.garimella@aalto.fi).