Skip to content

Latest commit

 

History

History
49 lines (37 loc) · 2.62 KB

Purple Feed Identifying High Consensus News Posts on social media.md

File metadata and controls

49 lines (37 loc) · 2.62 KB

#Purple Feed: Identifying High Consensus News Posts on Social Media

Key ideas

Introduction

  • Users play a role to select their sources for information as opposed to the past.
  • Many users limit themselves to news stories that reinforce their views.
  • This leads to a more politically fragmented, less cohesive society.
  • There are systems to promote blue information for red users and viceversa but users do not like them.
  • Purple news are news with high consensus of reader reactions. Hopefully these evoke a more unified response across society which might help to understand each other.

Contributions

Consensus definition and measurement

  • "high consensus if there is a general agreement in readers' reaction despite readers' political leaning"
  • consensus = abs((democrats disagree/number of democrats) - (republicans disagree/number of republicans))
  • Mechanical turk used to measure consensus from WSJ blue/red feed and Twitter top 10 publishers.
  • Confirmed that political extremes tend to pick lower consensus content.

Study of consensus of news post

  • Count Retweets to measure popularity. High consensus (158) and low consensus (177) on avg. Similar engagement.
  • Do they cover similar or different topics? Consensus seems to cover non-US centric political topics.
  • High cross-cutting exposure: # opposite leaning retweeters > baseline of opposite leaning retwitters for a publisher
  • Clearly high consensus correlates with high cross cutitng exposure

Identifying high and low consensus news

  • Features found

    • followers - on avg 67% are of the same political leaning
    • retweeters - active supporters, 78% of the same politicall leaning
    • repliers - 35% are of opposite political leaning
  • Use Kulshrestha et al 2017 to measure political leaning in [-1, 1]

Experimental evaluation

  • Supervised learning using ground truth from the AMt dataset. Used SVM, Naive Bayes, Logistic Regression and Random forest.
  • Pruned insignificant features and used 5-fold cross validation.
  • Accuracies near 60/70% for predicting consensus.
  • Publisher-based features are better than tweet-based features. Poor performance due to short size of tweet.
  • 74% of high consensus posts identified, 70% of low consensus.

Conclusion

  • We propose a complementary approach to inject diversity in users news consumption.
  • First attempt at operationalizing consensus of news posts on social media.
  • Deployed 'purple feed', a system to highlight high consensus posts.
  • Possibilities to study further evaluating the impact of showing high consensus posts on social media.