際際滷

際際滷Share a Scribd company logo

Media, Algorithms and the Filter Bubble

Geetu Ambwani
Geetu Ambwani

Featured Presentation: Critical Approaches to Data Science & Machine Learning Codes & Modes Symposium, March 17th, 2017. (http://ima-mfa.hunter.cuny.edu/reframe/)

1 of 24
Downloaded 13 times
Media, Algorithms and the Filter Bubble Codes & Modes Geetu Ambwani 3/17/2017 @geetuji
Filter Bubble Personalized algorithms serve up information that a user wants to see based on their likes and past history Users become separated from contradictory information resulting in isolation into cultural or ideological bubbles
Filter Bubble in News 6 in 10 Americans get their news from social media [Pew 2016] Confirmation Bias Personalization
News Breaks... News is now a choose your own adventure story Jim Rutenbergs NYT column 3/12/2017
A veer to the right
The boring middle
A veer to the left ...
Information Dissemination Through Social Media Retweet graphs of trending news topics [Garimella et al. WSDM 2016]
How should media respond ? We can design interfaces that help people read more balanced news in their environment of choice We can create compelling user experiences that nudge users to break out of the bubble
How should media respond ? 1. Show the opposing view 2. Show people their bias 3. Show source credibility
Show the opposing view 1. People say they seek diversity [Stromer-Galley 2003] 2. People agree with the norm of diverse news exposure [Garrett & Resnick 2011]
Show the opposing view
However ... 1. Recommending opposite views had a negative emotional effect [Graells-Garrido et al 2013] 2. people would still preferentially select information that reinforced their existing attitudes [Liao et al 2013] 3. backfire effect in which corrections actually increase misperceptions [Nyhan et al 2010]
Show people their bias 1. showed users feedback about their political lean led to a modest move toward balanced exposure [Munson et al 2013]
Show people their bias
Show source credibility 1. explicit knowledge of the source credibility or expertise not only affects what users read, but also how credible they perceive the documents to be [Vydiswaran et al 2012] 2. providing not only the valence (pro or con) but also the magnitude (moderate or extreme) of information source position is useful for encouraging exposure to diverse information [Liao et al 2014]
Show source credibility 1. explicit knowledge of the source credibility or expertise not only affects what users read, but also how credible they perceive the documents to be [Vydiswaran et al 2012] 2. providing not only the valence (pro or con) but also the magnitude (moderate or extreme) of information source position is useful for encouraging exposure to diverse information [Liao et al 2014]
How did we get here? Think incentives! The business of media is in dire straits Business Model: Ad $$$ in exchange for engagement Metrics that capture engagement: page views, unique users, session time
Newsroom Production versus Audience Consumption
How did we get here? Think incentives! The business of media is in dire straits Business Model: Ad $$$ in exchange for engagement Metrics that capture engagement: page views, unique users, session time Publishers and platforms are incentivised to optimize these metrics, to ensure their very existence Optimization
Where do the algorithms come in ? Optimization lies at the heart of machine learning. We model our problem as: Objective function to maximize System parameters that we can control Constraints that are limits on the inputs We learn from the data the parameters that maximize our objective function while satisfying the constraints.
Where do the algorithms come in ? Personalization algorithms all try to maximize engagement (revenue) given users past interactions. Unsurprisingly, some parameters learned are - Nuance does not click well @juliabeizer Some things that do click/share well are: Clickbaity headlines Kittens Articles that match the users confirmation bias
Where do the algorithms come in ? We can reframe our optimization problem as maximize engagement (revenue) given users past interactions with additional constraints The additional constraints could be to ensure a diverse mix of articles or to ensure threshold level of credible sources. Adding constraints revenue tradeoff
Conclusion To address the filter bubble, we need a hybrid approach of a compelling user interface and refined personalization algorithms. Always, the revenue tradeoff remains. These are questions that go way beyond whether we can develop AI technology that solves the problem, .So the technology exists or can be developed, but ... does it make sense to deploy it. [Yann Le Cun, Head of AI, Facebook]

More Related Content

Media, Algorithms and the Filter Bubble

  • 1. Media, Algorithms and the Filter Bubble Codes & Modes Geetu Ambwani 3/17/2017 @geetuji
  • 2. Filter Bubble Personalized algorithms serve up information that a user wants to see based on their likes and past history Users become separated from contradictory information resulting in isolation into cultural or ideological bubbles
  • 3. Filter Bubble in News 6 in 10 Americans get their news from social media [Pew 2016] Confirmation Bias Personalization
  • 4. News Breaks... News is now a choose your own adventure story Jim Rutenbergs NYT column 3/12/2017
  • 5. A veer to the right
  • 7. A veer to the left ...
  • 8. Information Dissemination Through Social Media Retweet graphs of trending news topics [Garimella et al. WSDM 2016]
  • 9. How should media respond ? We can design interfaces that help people read more balanced news in their environment of choice We can create compelling user experiences that nudge users to break out of the bubble
  • 10. How should media respond ? 1. Show the opposing view 2. Show people their bias 3. Show source credibility
  • 11. Show the opposing view 1. People say they seek diversity [Stromer-Galley 2003] 2. People agree with the norm of diverse news exposure [Garrett & Resnick 2011]
  • 13. However ... 1. Recommending opposite views had a negative emotional effect [Graells-Garrido et al 2013] 2. people would still preferentially select information that reinforced their existing attitudes [Liao et al 2013] 3. backfire effect in which corrections actually increase misperceptions [Nyhan et al 2010]
  • 14. Show people their bias 1. showed users feedback about their political lean led to a modest move toward balanced exposure [Munson et al 2013]
  • 16. Show source credibility 1. explicit knowledge of the source credibility or expertise not only affects what users read, but also how credible they perceive the documents to be [Vydiswaran et al 2012] 2. providing not only the valence (pro or con) but also the magnitude (moderate or extreme) of information source position is useful for encouraging exposure to diverse information [Liao et al 2014]
  • 17. Show source credibility 1. explicit knowledge of the source credibility or expertise not only affects what users read, but also how credible they perceive the documents to be [Vydiswaran et al 2012] 2. providing not only the valence (pro or con) but also the magnitude (moderate or extreme) of information source position is useful for encouraging exposure to diverse information [Liao et al 2014]
  • 18. How did we get here? Think incentives! The business of media is in dire straits Business Model: Ad $$$ in exchange for engagement Metrics that capture engagement: page views, unique users, session time
  • 19. Newsroom Production versus Audience Consumption
  • 20. How did we get here? Think incentives! The business of media is in dire straits Business Model: Ad $$$ in exchange for engagement Metrics that capture engagement: page views, unique users, session time Publishers and platforms are incentivised to optimize these metrics, to ensure their very existence Optimization
  • 21. Where do the algorithms come in ? Optimization lies at the heart of machine learning. We model our problem as: Objective function to maximize System parameters that we can control Constraints that are limits on the inputs We learn from the data the parameters that maximize our objective function while satisfying the constraints.
  • 22. Where do the algorithms come in ? Personalization algorithms all try to maximize engagement (revenue) given users past interactions. Unsurprisingly, some parameters learned are - Nuance does not click well @juliabeizer Some things that do click/share well are: Clickbaity headlines Kittens Articles that match the users confirmation bias
  • 23. Where do the algorithms come in ? We can reframe our optimization problem as maximize engagement (revenue) given users past interactions with additional constraints The additional constraints could be to ensure a diverse mix of articles or to ensure threshold level of credible sources. Adding constraints revenue tradeoff
  • 24. Conclusion To address the filter bubble, we need a hybrid approach of a compelling user interface and refined personalization algorithms. Always, the revenue tradeoff remains. These are questions that go way beyond whether we can develop AI technology that solves the problem, .So the technology exists or can be developed, but ... does it make sense to deploy it. [Yann Le Cun, Head of AI, Facebook]