POLINA:  Okay.  Can everyone hear me?  Good?  No.  All right.  
I'm just going to hold it.  All right.  Can everyone hear me 
now?  Hello, everyone, I'm Polina Giralt.  I'm a data engineer.
ANITA:  Hi, I'm Anita, and I'm a data scientist.
POLINA:  And we're super psyched to be here to talk to you and 
Measuring Virality at BuzzFeed.  So.
ANITA:  This is what we're hoping to cover today.  We're going 
to give you an overview of BuzzFeed in case nobody has heard of 
us.  I like the reaction.  I'm going to tell you a crazy story 
about a dress and Anita is going to show you some amazing graph 
data visualizings that we've come up with, we're going to talk 
to you about challenges and key takeaways.
ANITA:  Okay.  Who are we?  So BuzzFeed is a tech media company 
that was founded in 2006 as an RND lab to understand how content 
goes viral.  Today, we're home to all kinds of addictive 
content, so cat lists, memes, investigative journal, and long 
form video.  As of earlier this year we see about 200 million 
unique views per month and we have 900 employees around the 
world.  So when Polina and I tell people that we work another 
BuzzFeed, a common question that we get is this.  It's kind of 
cute.  It's also kind of true.  We do care about page views and 
clicks.  That's not all.  We're much more interested in shares.  
Specifically although it's practical to trick people into 
clicking -- think clickbait, it's way more difficult to trick 
people into sharing.  It's really difficult to force a meme.  
What's cool about BuzzFeed is that over 75% of our traffic comes 
from share referrals.  Great.  So we want to encourage users to 
share.  But how do we measure a share?  Hold onto that question, 
we're going to come back to that in a second, but Polina is 
going to give you a real-life example.
POLINA:  Okay.  Cool.  So is there anyone here who's never seen 
this picture?  Anyone?  Anyone?  Seriously?  Leo, what color is 
this dress?
LEO:  It's white and gold.
POLINA:  Right answer.  Who else sees white and gold?  All 
right.  Who sees blue and black?  All right.  Does anyone see 
other colors?  Okay.  So on February 26th, at 6:00 p.m. on the 
east coast, Kate published a heated question and asked readers 
to settle this, that simple question becomes a singular moment 
in global culture, the largest traffic day ever for BuzzFeed.  
Celebrities debated it publicly, usually on Twitter and couples 
and friends argued right on their mobile phones.  By 8:30...
[ Laughter ]
By 8:30, alerts were going off in a dedicated Flack channel.  
But sysops was on it.  The gif on the right shows what peak and 
offpeak traffic looks like at BuzzFeed.  On the left, this is 
Google analytics and from each address at 10:00 p.m. showing 
over 670,000 concurrent users.  That's at the same time.  By 
10:30 we've got a couple of posts that are getting backed up 
from publishing but half an hour later, we've published them and 
by the end of this, we've actually added 40% more capacity to 
our servers.
           So we definitely have a culture of when something's 
blowing up, we swarm it.  So as a result, we quickly translated 
this into five different languages and in less than 24 hours 
people from all -- every corner of the world were looking at 
each other's phones about this dress, you know, trying to figure 
out what color is it.  And by the way, this really awesome map, 
Anita made that.  And our record-breaking win was that readers 
never really had any technical problems at all either looking at 
the post or taking that poll.  Okay, so next, Anita's going to 
talk about graphs and what kind of statistics this data looks at.
ANITA:  Cool.  Thanks, Polina.  Sorry.  Guys.  I'm having issues 
here.  Here we go.  Here we go.  Sorry -- sorry about that.  So 
let's talk about data.  So this is an example of some 
traditional web analytics data.  Say I share a piece of content 
and I'm interested in what social media platforms it's been 
shared on.  And I'm also interested in, say, the time of day 
during which those shares occurred.  This provides us with a 
really limited view of the social web.  This is an example of 
the type of data that we're working out of at BuzzFeed.  So a 
graph.  Suppose, for example, I share an article about Taylor 
Swift on Twitter and one of my followers decides to pick it up 
and decides to share it on Facebook.  That information, that 
share is captured by this directed arrow.  So some quick graph 
theory 101?  The circles represent nodes of the graph, those are 
views.  And the arrows are the edges, so all of a sudden we have 
the ability to study things like diffusion and propagation to 
answer old questions in new ways and to answer new questions 
altogether but what does this look like in practice?  This is an 
example of a cascade, a process that shows information being 
successfully passed on.  In this case, a piece of information 
refers to a view.  The cascade is rooted at our original 
Facebook post, "rooted" meaning that the original head node is 
at the Facebook post.  This cascade shows 8.6 million views for 
the dress.  This is just one cascade of many.  So what exactly 
is this graph showing us?  Let's look at another graph, it's a 
little bit smaller but we can interpret the colors more easily.  
So this one's rooted at our official BuzzFeed Twitter account, 
it has about 975,000 views and the colors of the nodes 
correspond to the time in which the view occurred, so gray is 
dark social, so chat, hale, and white a blogs and other 
publications.  The green corresponds to the edges, i.e., the 
shares between the nodes.  But we can zoom in.  So let's say we 
zoom in on this subcascade.  This is potentially a journalist 
who I don't know, picked up our link on Twitter and shared his 
or her publication.  This is a closeup of it from a couple of 
different angles and we can continue to zoom in.  Notice that 
the nodes mostly appear to be light blue.  Zoom in on that and 
we can zoom in some more.
           So the cool thing about this is that possibly, for 
the first time in history we have the ability to see millions of 
shares across multiple social media platforms.  That's pretty 
cool.  Okay.  So now, I'm going to hand it off to Polina to talk 
about the technology that's underlying this.
POLINA:  All right.  Cool.  So those are really totally sweet 
graphs that Anita just showed just to emphasize that.  But how 
did our team measure those shares?  So one of the projects that 
we get to work on is called Pound.  It is a backronym.  It 
actually stands for the hashtag, obviously.  Pound allows us to 
collect and aggregate data so that we can study their diffusion 
and content propagation across the social web with the pound 
data, our data scientists can measure, study visualize virality, 
and give teams gain insights into their work.  Of course, like 
any good project it comes with a share of technical challenges.  
So as we've been designing and iterating and architecturing 
this, we've changed a whole bunch of stuff.  We've gone from 
processing logs with MQ to real-time NSQ data streams.  We added 
a redis caching layers everywhere.  Data obfuscation because 
privacy is super important.  All the data is anonymized and 
aggregated and there's no personal identifiable information 
whatsoever and we're always validating the data that we're 
collecting which I think is one of the hardest problems in 
computer science.  So what's next for us?  We want to continue 
building our tools because we want to keep showing our viewers 
larger and wider range of content and we do this by machine 
learning and other cool data science projects.  And since 
sharing is caring, we're also sharing our research with 
researchers who are studying networks.  And you can check our 
blog for more information on what we've been up to.
ANITA:  So if there are four main points that you should take 
away from this presentation it's the following:  We optimize 
sharing.  The dress is an example of a piece of content that 
went viral and we have the ability and the technology, 
especially using pound to capture graph data that enables us to 
answer old questions in new ways and to solve new problems 
altogether.  But most importantly testing and learning is key to 
everything we do.  Essentially, we iterate and we iterate a lot. 
 The analysis that our data scientists use inform engineering 
decisions, which in turn allows us to do even more analysis.  
And before we conclude, we just wanted to give a huge shout-out 
to our amazing team without whom none of this would have been 
possible.  And that's pretty much it.  Once again, this is 
Polina, I'm Anita.  Yay, data!
POLINA:  And that's data!
[ Applause ]