This is a liveblog of Lada Adamic's plenary keynote from Political Networks 2014.
When your friends deliver the newsLada Adamic is a Data Scientist at Facebook and former associate professor at the University of Michigan's School of Information. Her talk is entitled "When your friends deliver the news." Using studies based on Facebook data, she invites us to think about factors of social networks that affect the spread of information.
She opens with a set of questions and concerns raised by the idea of your talk's title: what happens when your friends deliver the news through what they share on Facebook...
- How is your exposure affected? (Your friends are not a random sample of the population nor are they mainstream media journalists.)
- Does it affect political engagement? (How interested you are or likely to vote.)
- What do social movements look like? How does success propagate?
- How does any information spread, is it predictable?
- Is it reliable?
Lada shows us the top five most shared stories from a year ago (i.e. June 2013):
- Drowning doesn't look like drowning (Slate)
- Boy's death highlights a hidden danger: Dry drowning (Today Show piece containing substantial misinformation)
- 22 Maps that Show how american speak English totally different from one other (Business Insider maps that were later integrated in the most read NYT story of 2013)
- Edward Snowden: the whistleblower behind the NSA surveillance revelations (The Guardian)
- 8 Foods we eat in the us that are banned in other countries (Buzzfeed)
She tells us that women in their 40s read the drowning pieces, whereas men in their 20s read the Snowden piece. There are definitely clusters of people more likely to read certain stories. But Lada asks: Is there a filter bubble? Do we get echo chambers, especially across political lines?
Using information from users' profile pages, Lada and team members rated what people's political leanings were: very liberal, moderately liberal, moderately conservative, very conservative. And they coded the different news sites by what was read by users of certain political leanings: from ThinkProgress on the far Left to FoxNews on the far right. They found that content skews liberal in aggregate over Facebook even though ideology is balanced. There is simply more liberal content shared by all users on the network. Lada cites Duncan Watts' Friend Sense app research, and asks: Can we understand what the egonetwork of a conservative looks like?
People's friends aren't exclusive to their political beliefs. They found that the distribution of friendships skews toward their own political leaning but still retains some balance across the spectrum. Then they looked at the interactive patterns between users and news across the political spectrum, breaking them down into four buckets of user-news interaction:
- Potential: all of the news your friends are sharing
- Exposed: what showed up in your news feed (balanced diet of liberal and conservative for conservative users)
- Selected: what was clicked on (no effect, they were clicked on in proportion to what showed up in feed)
- Endorsed: what is liked this is where the difference exists, conservatives are much less likely to endorse liberal news
They found that endorsement was the key difference between users of conservative versus liberal ideology. Conservative users were significantly less likely to endorse liberal news, even though they were served and even clicked to read liberal news at similar rates to the liberal users.
Research by Solomon Messing and Eytan Bakshy looked at and experimented with activity on Facebook prior to the 2012 election. The treatment involved adding more political news in certain users' news feeds. They found that people reported being more interested in politics and government when, unbenkownst to them, they were getting more news in their feed. There was a greater effect for users that don't log in everyday, since people who log in everyday are more likely to read everything. The treatment group also reported they were more likely to vote, with a stronger effect again on less regular users.
What about active political campaigns?Eytan Bakshy studied the HRC "equals sign" campaign on Facebook. One to two million people participated in the campaign. We see that profile updates really did surge when the campaign kicked off. Lada suggests that homophily might be involved. And it might require something like seven friends doing it before you do it, i.e. it works like a "complex contagion."
In the case of James Fowler's research election year project (Bond et al. 2012a>), the treatment involved showing the faces of people who had voted versus just aggregate giving aggregate numbers. When they broke it down by voting district, there was a boost in the treatment locales, but it was complicated by how strong the tie was. Stronger ties had more impact.
How exactly does information spread in social networks?Lada asks, What are the factors that affect the spread?
- network structure, e.g. who initially posted it?
- temporal patterns, e.g. how quickly it propagated?
- content?
- origins?
- veracity, e.g. was it true or not?
And if we understand these factors can we predict large cascades?
This was Justin Cheng's summer internship work at Facebook in collaboration with Alex Dow, Jon Kleinberg, and Jure Leskovic ("Can Cascades be Predicted?"). Of course, most things don't spread as Goel et al., 2012 tells us. In the case of Facebook images, 50% of all shares come from just 0.5% of the top 5% of images reshared.
Lada gives us two different examples of heavily reshared images:
Obama's Victory Hug gets 600,000 reshares
"Cathrine says that if I can get 1M likes, she will have sex with me" gets 100,000 reshares
You expect the Obama Victory Hug image to go far because it's coming from the highly followed Obama Campaign page, but you can't predict that based on the follower numbers alone.
Lada offers the exercise: Given a cascade that currently has size k, predict whether it will reach the median size f(k). This is a useful approach because half will cascade and half won't at size k. Using this framework, she assesses the importants of the above factors against modeled cascades.
When looking at original feature importance versus k, she finds that views are important initially but less over long term, origin from a page has little effect especially initially, and how many people reshared from the root was initially important but less so over time.
Looking at resharer feature importance versus k, she finds an interesting demographic trend: older people reshare more, young people create more original content. When it comes to content feature importance versus k, captions are important initially, but not over the long term; synthetic content and overlaidtext are both important though, i.e. meme-like content goes farther.
In assessing structural feature importance versus k, average depth plays a role; there is a trend whereby things that are more viral tend to do better. There is definitely some rich get richer in cascades. But the temporal features are the most important. Timing really matters on cascade size.
But predicting size isn't everything says Lada, can we predict structure? She says, yes you can predict shape if you observe it for a while; if you see it going deep it will keep going deep.
How do rumors propagate on Facebook?(See Rumor Cascades by Adrian Friggeri, Lada Adamic, Dean Eckles, and Justin Cheng.)
Facebook is a neutral platform, within its Terms of Service, but you can reach a lot of your friends at once. And Lada argues that we tend to trust the information that reaches us via our friends, which makes the medium a great substrate for rumors.
She gives the example of the Cabela sporting goods receipt rumor, where Cabela accidentally applied a medical excise tax to gun equipment. It becomes a meme through its caption, which presents it as "evidence" of what Obamacare is doing. The receipt is real, but of course the caption is misleading.
Lada shows us the rumor spread as an information cascade visualization in which they have chained together clicks and shares to show how it spread.
What's unusual about this rumor's reshare pattern is the fact that its chains are really long: going from friend to friend to friend. Most reshares don't achieve much depth, she says, but for rumors you get longer chains, regardless of the veracity of the rumor.
The Facebook Data Science team used links to articles on Snopes in the comments of reshared rumors to identify them for analysis. They found 16,672 cascades, collectively reshared 62 million times. They looked at the entire Snopes corpus online and found that Politics rumor pages were most popular followed by Fauxtos, and Inboxer Rebellion (chain mails propogating as images). Uploading political rumors is very popular on Facebook, but politics doesn't propagate well. Rather, when they looked at the cascades, it was actually the email chain letters (already perfectly designed for sharing) that really spread well on Facebook. Additionally, Glurge was really popular (too good to be true feel-good stories)—Rose the 87-year-old college student being particularly popular.
False rumors get more comments linking to Snopes than the true ones do. But what happens when a reshare gets Snoped? Performing a term frequency analysis on comments shows that words like Urban Legend, Snopes, Bullshit, and Fake are associated with false rumors, where as Amazing and Respect are associated with true rumors. Lada said they also found that cascades continue even after they are Snoped in the comments, but there is a higher chance that rumor post will be deleted if Snope is in comments.
Are rumors constant? Lada sees rumors that have multiple variants, like the "moneybags" rumor. This one pops up every year with the numbers changed, sometimes correctly and sometimes incorrectly for the year in question; it's easily manipulated and reposted. Lada says sequences of numbers and letters will stay the same in these types of rumor variants. Amber Alert rumors will often keep the same license plate and change the car make or the city.
Can we counter rumors? Lada notes that sometimes a humorous variant will take over the original. An example is the "Facebook is going to start charging rumor." This one also recurs but was recently overtaken by a parody in the style of an overlaid text meme. This copy and paste meme was more effective than all the fact-checking comments that responded to false rumor.
ConclusionsLada draws three conclusions for the Political Networks audience from her data science work at Facebook:
- Social networks play an important role not only in disseminating but also in shaping information
- Political and other news are increasingly being distributed via online social networks such as Facebook
- This has important implications for interactions and interest in politics
Question & AnswerQ: Does this look the same on Twitter?A: Twitter cascades tend to be a bit shallower. Goel et al. 2012 showed no large cascades on Twitter, rather it was more dominated by large nodes. There is also less of presence of Snopes on Twitter. There are interesting rumors on Twitter but they don't have the same dynamic. The news outlets are on and active on Twitter, and they often tend to be the source of cascades there.
Q: Which jumps on stories faster, Twitter or Facebook?A: They both move fast. And they will both generate large cascades if say a celebrity posts something that starts taking off on both their Facebook and Twitter pages.
Q: Are there boundary effects on cascades, such as language barriers and geographic origins? Second question: Did you ever see rumors just get immediately shut down?A: Language barrier is definitely there. She did look at a rumor that got translated into other languages and those clustered by language then. Sometimes different languages have cultural differences in the way people comment on them. Faloutsos has studied the rumors online showing that the decline drops off quickly like a power law. They looked at one rumor about a Japanese power plant was spewing toxic stuff into the air, but when the power plant posted information
Q: I wouldn't have guessed we would have had this rich data ten years ago. What are people imagining about the kinds of new data sources we'll see in the next ten years?A: That's a hard question to answer. As more and more activity moves toward platforms like Facebook, things will get more interesting. For instance, what if Facebook becomes the dominant platform whereby you get to know your politicians and engage in debate and now that is all in structured data? We could start connecting conversations to emails to congressman or other kinds of impact.
Q: Interested in your finding about how temporal characteristics were more important than structural effects. Do you see evidence in the data of some type of initial catalyst like an "event" as would be part of traditional PR/marketing.A: I didn't mean to discount structure altogether, they are tied closely together. In the case of the Obama Victory Hug, when Obama posts there is a huge spike, then another spike when Michelle Obama's page reshares it, and a third when Alicia Keys shares it. These are important events. But in other memes, like the 1M Like photo, that one was more organic rather than involving one or two key figures or a singular event.
facebookdata scienceinformation cascadesrumorsmemessocial networks
