Okay, well, hello, here I am with Igor
[FOREIGN] and Sam Shaw at LinkedIn
headquarters at Mountain View.
[laugh] And, what I would like to ask them
is how they first got into social network
analysis.
Igor already started this story which
predates LinkedIn.
So, I will let him continue.
So, I'll just tell it again.
So in the beginning when I started within
social network was 2003.
So that is really like in the dark ages.
At that time the field was more driven by
consultants.
And you had these views of small networks.
I was playing in the role of Enterprise
knowledge management and Enterprise
knowledge management was the search for
document at manner for something that it
has I need to do and for expense to help
me out and the koffing that we had is that
within each document we could create a
symetic math that you should come the
document back that I give a account that
counts these concept that let.
So its a nice reservation right.
On the other side we have names of people.
There are experts, and we felt like that
these are kind of static.
I mean, I don't know who they are,
because, otherwise, I would have directly
reached out directly to them.
I don't know how they're related to one.
So we took that and put it in Cindy's
plate.
Because we had interactions around
documents.
We have the shares that we have today.
So we could figure out why these two girls
or these two ladies were part of the
expert that we will treat back.
And then it created a map.
It created one of these social maps.
So that, for me, was my a-ha moment.
Well, because, small start up,
40, 50 people.
You read the map and say hey that's right,
that's exactly how we work.
It did work.
Mm-hm.
I mean that was a small group set up 40-50
the company make it so we stop doing this.
[laugh] Now it's way different, now its
maps of hundreds of millions of users but
there the difference is you get to play
with real society.
And you can actually see how it's alive,
how it changes.
That is phenomenal.
Mm-hm.
So, so one thing I've been telling
students is how it's different when you
slice a small network.
And people can easily recognize oh yes
these are actually the groups that work
together and so on.
And we'd imagine that with a large network
you have a lot of overlaps from
individuals changing companies, that
doesn't make the same amount of sense.
Or do you need to slice it before you can
really work with it?
I think Sam and I can talk at length about
that. How many pixels are there on the
screen?
Sort of less than one hundred eighty
million depending on the resolution.
See if I put a dot to remember.
I put a dot remember on Twitter, on
Facebook and on LinkedIn.
Well we have this screen that's full.
So you have to slice and dice it to some
extent.
And through that, you start getting
actually the nice thing. huh Yeah.
Many, a lot of challenges comes in how you
are going to slice it too.
Because, and as we have done, run into
these problems where, slice it in a
certain manner then you lose the
information that you want to look at.
And you slice it in a different manner.
So, you look at LinkedIn in the Bay Area
performs very differently than it does say
in easily.
And, and looking at that is, is And
figuring out how you are going to slice it
is quite a functional option.
Uh-huh.
It's also leaked to me, he's back on his
steps.
And in the old days it was like the myth
of the average man.
There, there's an average person as an
ever did.
Well, the average person is half female,
half male.
The average family is 1.5 kids.
That doesn't exist.
The slice and dicing makes me reminded is
not an average user.
Mm-hm.
There's not an average figure to look at.
The different aspect is.
There's, there are the students.
There are job seekers.
There are professionals.
There are this, there are that, there are
that, that.
Facebook, you can have there are
teenagers.
There are early teens.
There are late teens.
They're all different.
[laugh] No, no, no.
Definitely not.
Definitely not.
Actually one thing that I noticed is that
it's drastically changing.
I don't know if, did we meet in 2005 at
that Microsoft meeting on social
nzetworks?
I'm not sure.
They brought bunch of research, Research
material for their dance. And they brought
kids, 13-14 year olds.
And they start saying, I do this on my IM,
I do this on MySpace, I do this on the
phone.
I do this and that.
And we were all flabbergasted.
What?
You do that on this?
And it goes like, you don't break.
Don't break up with someone on IM.
You called him up or you don't do this on
the phone.
I forgot what it was.
These individuals are our users.
And, we all benefit from the perception of
prevalence of social networks.
That it's not something, I don't know what
to do with it.
It's almost inbred into their DNA.
Where as my generation, we going to look
at my sister.
Why this, why that.
But my dad, pffft.
But that changes drastically, so it's not
a,
They, have the whole ecosystem sort of
adapted to it.
Mm-hm.
What we will be in five years.
I don't know, but it becomes so much
easier to do on it, that what we actually
see.
What they want from it.
Much easier than us telling them what it
should be look at.
Mm-hm.
Mm-hm.
And I, I guess there's kind of a, a two,
two part question.
To what extent is individual speaking
here, when they use a network such as
lifting different from holding, or the
same as what they might be doing when
they're doing regular networking or
hanging out?
And also to what extent you know, such
behavior might inform the algorithms that
you use in trying to facilitate those
interactions.
Do you, do you try to simulate what might
happen, and?
And off a.
Mm hm.
Offline.
Mm hm.
Is actually.
[laugh] Yeah, I mean, I think there's
quite a...
I mean LinkedIn is a little bit different
in that it's, you know, I mean your
information is inherently public which is
like, you know, you're putting your CV out
there; you want to be found, is what our
core kind of principles of having this
identity on the web.
So, like you know, when you have that kind
of dynamic things change: like I think
people are much more conscious about what
they want to present and then they change
their behavior, I think, and...
To kind of reflect that, which you would
do differently, say in an offline setting.
And we've, I-, it's, it's, I know, it's
hard to actually understand what would we
do.
Like the differences between that and kind
of.
Of whatever that we've actually.
Have a really true chris butter sandal.
But what about, not right now as we are
recording, but in two weeks.
The students will be learning about
Milgrams experiment and passing messages
six times.
And to.
You know, it, I, I think it doesn't may be
happen naturally, but, it, it, but it
could, would be place where you might
facilitate, people actually navigate.
Mm, mm.
The.
Mm-hm.
The network.
To what extend do you, do you see, and
what kinds of smarts do you?
Things just kind of blow up.
Right?
The number of friends of friends of
friends.
Mm-hm.
The.
Can we?
The.
Again, 2003 there were three start ups in
social networks.
Friendster, Spoke and LinkedIn.
If you look at Spoke.
Spoke was a city dealing in business
transactions.
I am a salesperson, I want to sell to
company X, I don't know anybody in company
X.
But who do I know that knows someone?
It's kind of similar to Stanley's
experiment, like, how do I get my message
to that person.
Well, I don't know who that person is.
In this case, I don't know who that person
is.
You have to set up who you are trying to
target.
I always felt that I would be the most
successful start up.
In that domain, of course, you haven't.
[laugh] I was wrong.
I'm fine with it.
We.
We have a, a couple of parts that we're
looking into.
One the pathfinder which is actually
trying to surkee, that this is how we are
interconnected between a and b.
So, making that path explicit and making
you the choices of which path do you want
to go down in order to convey a message or
do anything across the world.
And it's, it's slightly tricky, because
you have multiple hops to go through,
multiple types of connections.
Whether you're directly connected to the
person, whether you share a group with
that person, whether you live in the same
area.
All of these sort of interplay differently
within the algorithm, and we're still
learning how to actually make that think
better.
Huh, huh.
So, so any, any previews.
[laugh] Nope unfortunately.
[laugh] No, no?
Okay That's one of these things, these
cool things when you work in places like
that, did with something before.
We do have the data to try things out.
Most of our ideas don't make it out.
Because it's, either it's, it's extremely
hard, or the, the signal is not there, or
the iterations.
And we, it's gonna take longer than we
thought, and some do.
And when those do, sometimes they are
extremely surprising to us.
Mm-hm.
We just launched two product.
Yes.
Which we, we laugh about it every day now.
Which are extremely viral, and they behave
virally differently.
Mm-hm.
And to understand why is actually a very
fascinating question to think about.
So tell me more.
It sounds like you had something to do
with this.
Oh yeah.
So we did this thing called endorsements
where you can provide kind of a
lightweight, thumbs-up sort to basically
on your colleagues' skills and expertise.
And we really under.
It is really interesting to see that kind
a social dynamic in that future it can
happen as people slowly get, get to
normal, go to quality figure out skills
they have.
It actually go and see this kind or
egoistically got start, so I'm going to
give them that and what is that actually
change ego's perception, how would it
actually spreads that thing in entire
network.
And I, that, that, that thing has been
growing immensely, much more than we
actually even anticipated.
And people are really in tune to that kind
of thing.
So where do you get the expertise?
Oh yeah, yeah.
Only for you.
[laugh] Off the record.
Only for people watching this, no one else
is gonna know.
So you can do a few things, you can take a
look at profile information, and be like
okay, you know, based on the descriptions
to be like hey What do my friends know?
What do my connections know?
Mm-hm.
That can...
Mm-hm.
...
Basically drive what I know.
And then you could also do the same thing
about your company, your title.
And you bring all those pieces of
information together.
And you could have a pretty, or a decent
rep-, representation of what, some skills
someone might know.
Right?
And it doesn't have to be perfect.
But then, you know, you present is as a
user.
And then they kind of define that.
And it's been wildly successful, using
that kind of approach.
Mm hm, mm hm.
Cool.
So kind of friend sourcing, expertise.
Yeah, sort of...
It's sort of, like, playing on the
Monopoly thing.
Yeah.
Like, you've got people with that skill
things and sort of.
Clustered together on top of it were
professional networks that you would have
those connections.
Right.
And you let them.
I just think that, that source of data.
Huh.
Huh.
So, I have done a little bit of research
of what people say publicly and what they
say privately?
[laugh].
Have you thought about that at all?
Publicly I say that Igor is an expert.
[laugh] And privately?
[laugh] It goes back to one other question
that you had earlier.
The, The difference is that everything
that you do on Linkdin is, is tailored.
I mean, it, it has, it has your voice to
it and you're identified through it.
You're not anonymous.
Mm-hm.
You're not using a, a weird kind of name
on that.
Mm-hm.
It's really you with your name and that I
think effects what, what knowledge goes
through the network.
Mm-hm.
It effects how you express it and it
effects how you are not expressing some
things.
Which makes it very easy to sort of in
front of it. At least for you guys you
might be able to make some of princes when
people don't endorse, [laugh], [laugh]
that would be a Yes, we cou, I would
think, I would care, I would think about a
couple of odd ones.
What?
The, the really big difference is that On
your profile, you get full control about
what it is out there.
So if I'm saying you're not good at
something well, you get control whether
you want to show it or not which probably
we won't want.
But then on the other hand a lot of
targeting that we have or recommendations
could take that into consideration.
[laugh].
Okay so this is part one.
You said that there are two things that
went viral.
[laugh] The other one is, we just launched
a way for.
For having richer content on the site.
So normally you would have like a small
almost a network feed which is about 140
to a bit more characters.
And we give the opportunity to some
individuals to say, "Well, you are an
influential in this domain, why don't you
use a longer post?" Almost like a blog
type?
Uh-huh.
And again, it's standard, it's, it's, it's
keyed on, on your identity.
And we ask about 140, 150, influentials
to.
To message.
And these are, from one side, the, the
Romneys and the Obamas today.
To the Richard Branson, who's the most
popular one.
To, the head of the world bank, to, a
bunch of individuals that are.
Whose opinion you might want to, to read
from time to time, and you hope from them
to get the very good content part.
So that's another one that's not well.
You see, there's some individuals that we
follow this person, that person.
Um-hm.
To get access, to their, to their person.
Um-hm.
Or to get dynamically access to their
person because you call always go on the
public.
But to dynamically hear that to be
notified and that has a different way to
progress through the network.
Cause I follow that say you or him.
Then you say oh he's following him and her
and maybe I should check what he's saying.
Um-hm.
And, and that has a different way to
diffuse as compared to the previous one.
Okay.
Yeah.
Sounds neat.
So what's diffussing is the polymer.
Yes As opposed to this activity which you
are still curating.
You're Correct.
You're identifying.
One of the high level if you look at it,
the posting on an infant here.
It follows.
It's something new, it's the follower
posting.
The other one is the endorsement.
Is there is such a thing as an
endorsement?
What is it?
They knew it was before.
What are these skills thing, that I knew
what I had been, that was before?
They are both very, especially today, it's
been, what, about three weeks, four weeks
since we started endorsements?
Yeah, three weeks.
And by in front here is about a week.
So, today you are really seeing this.
This thing explode through the network.
Mm-hm.
Mm-hm.
And, I guess identifying influences here.
You mentioned, you know, maybe, maybe you
can use these endorsements?
And, and some fun to you actually.
[laugh].
Well. How does one do that?
[laugh].
As it always happens.
Uh-huh.
Sometimes we're wildly more successful
than we thought we were.
Mm-hm.
Going to be.
So we were rather careful to create very
well at the beginning.
Mm-hm.
So we went through, a high level of
editorial review.
So there's just going to be a few of them,
and there were a couple of things that
needed to happen.
One is that they were willing to actually,
do this porting, so that it seems dynamic,
and not that I'm not posting one every
year.
Mm.
And that you have this mechanism.
And then what we did is, well let's get
ready for that second wave.
If it works well, let's figure it out.
And there we used multiple signals to
figure out who is a potential influencer.
Within the network.
Part of the signals that we were using is
the skills map.
Part of the signal that we're using is
that people are sharing content at
LinkedIn, and what happens to that content
at issue.
And people share content within the
LinkedIn share button that's under where.
What happens to that content?
It sounds just being static, or if some
where else being picked up.
So they were already acting like in
trances without them really knowing that
they work because they're have the crowd
that would actually follow what they, what
they click on.
Mm hm.
What they acted upon.
Okay.
Ooh.
So, so Yeah we'll, we'll be getting to
the, I'm sorry, I smacked you.
[laugh].
So I remember you talking in Dublin some.
A few months ago now.
About predicting where peop-, where people
are going to move next for their next job.
And somehow also incorporating geography.
So...
Mm hm.
[laugh] Uh...
Should we sample this?
[laugh].
The interesting thing that we have done
these maps for the US. Now I'm invited to
go to Dublin, and say oh, I don't have the
maps for Europe And fortunately enough,
and even after the meeting, some of the
people who weren't in the audience.
Labor guru of something in Europe.
What we were doing is we were like well,
from your resume, I know where you're
working and where you are going to work.
And from that I can infer maps of are you
speaking to a certain region?
Are you willing to get a job opening to
another place?
If you look at it says well the migration
pattern in real life.
And we use part of that signal to figure
out which job I should recommend to you.
It's.
Mm-hm.
It's one of the things that I wish we
would do much more on Linkedin.
Much more looking into this labor market.
Uh-huh.
One thing that the reason we talk about
was the skill gap.
How come some people are not employed when
we have so many job opening?
Oh.
Is there a gap between the, the skills of
what you need for the job compared to what
you have.
Mm-hm.
Another part is well.
I thinking of taking some Coursera
classes, right?
[laugh] No.
That actually.
No seriously.
That's one of the things we are trying to
figure out.
Well if there is such a gap.
How do you,
Provide, that is organizations with
individuals.
See I', m looking for a job x, and there,
well there's not that many job x on the
market, but there's but this is job y on
the market which is close to the job x,
but these are the skills that you missed.
How do I get these guys?
Well I can work on this, but I cannot get
the job.
So I cannot work on this.
So how do I get it?
Well, classes, courses, books, online
this, online that.
Mm-hm.
And, and provide a way to actually get
that, or how to get internships.
Or maybe you can discover stud companies.
[laugh].
[laugh].
So, maybe you can't just directly go from
[CROSSTALK].
Say now, wait a minute!
I'm going to mark that down.
[laugh].
[laugh].
That,s.
Yeah, I think it's really interesting in
the context of veterans for example.
Like we have, you know, we were open up
these things where, you know, veterans
come they, they leave the military.
They almost have the skills but what
skills do those map to in, you know, in
non-military contacts?
And, you know, how, like what can a radio
operator in the military, what, what are
they really qualified to do and how can we
help them get a job which can more utilize
the type of skills they used in the
military?
That was actually really interesting kind
of thing to look at, and deep dive into.
Mm-hm.
Mm-hm.
So I guess there's still this question of,
the textbook social network analysis and
then social network analysis that scales
to the amount of data that you have and
you've said a little bit about it.
Is there anything more that the students
might want to brace themselves for?
You know?
[laugh].
[laugh].
Oh, yea.
I think almost everything you read in a
textbook will break down in scale.
We've noticed that quite a bit, I think.
I think a good, classical example of that
on LinkedIn is we have this people you
might know feature, which, you know, tries
to go and find, the relationships on the,
the social networking platform.
And you might think, hey, I'm just gonna
go and, like, look at second degree, like,
close ties and figure that out.
It turns out that really hard to do when
you have 185 million members.
And.
Actually doing these things at scale and
actually figuring out these well, these,
actually was very difficult.
And it was very humbling.
I mean, words like ten million.
It quickly breaks down when you get to 50
million or a 100 million things.
And you can think of different kind of
approaches to do some of this stuff.
Some of this can be parallelized, but you
know, a lot of this stuff actually cannot.
So you have to think maybe about
approximations or other things you can do.
Or really put domain knowledge into the
problem,
To really understand how you can get these
algorithms to scale.
You know, quadratic is not too really good
for us.
[laugh] Just to say that reading something
in a book.
I am going to a course to manage this.
[laugh] Yeah.
Yeah, I am not saying don't.
Yeah, the theory is really important to
understand.
To basically understand the applications
of it.
And then you know to really distill it and
deeper dive into these problems.
The way that I was viewing it is Being a
statistician, the background is a bit
math, so I started as a mathematician,
then a statistician.
Mm.
And I'm now dealing with money.
Mm-hm.
And there's a fundamental difference
between a, b and c.
Mm-hm.
The math is, is geared into, the
mathematical model.
Axioms, theorem, Lima.
Everything is true or false in some
extent.
At static fusion, nothing is true or
false.
Everything is uncertain.
But you still use math as the, the, the,
the tools that gets you towards, you know,
what is optimum, what isn't optimum.
And you understand the methodology.
Then you go to, billions of data, or
millions of this.
And suddenly, even doing a median is very
complicated.
And you go back to the average.
And you say, well, I know that the median
is better as a point is in with the
average, but, hey, you know what, I, I can
only do one average fast enough.
Mm-hm.
So I have to live with it.
But then, at that point of time, what you
do, in the back of your mind, you know
what is, what is a fundamental assumptions
of different techniques.
And you can use them and tweak them a
little bit.
So you know you are not doing the perfect
thing, but at scale, you have.
No options but you know what is the impact
of the change.
And you can maybe cry for that bias or
not.
Mm-hm.
Because their not on part time, but they
are just way too big and way too much.
Well, so, those were some of the
challenges.
Any, anything else that we may have
missed?
There is a beauty on scale.
It's not, we, we shouldn't be afraid of
scale, but, did it, it's really
fascinating, to see things that when they
have ton of traffic.
Because its, its makes it, it gives it its
dynamism, it gives it its diffusion
signature and because it had skill, the
other duties are I think its hard and
because its hard its interesting.
So I think its a chance that people should
pick up and up your favor.
Get tons of things that would fail but
it's a very, very fascinating thing to
play on.
Yeah, and the other thing is you can learn
different things at scale.
Things that, you know, with a small amount
of data becomes very difficult to
understand, but when you have a large
amount of data such as LinkedIn has, you
can see patterns that you would not
necessarily see.
Which is actually, I think a, very
interesting angle,
But. Mm-hm.
So really large things under the
microscope.
Yeah.
Yeah.
Exactly.
Yeah.
That's a good way to put it.
Yeah.
And the other thing, these are fascinating
times to play with that there.
Some years ago even if you thought about
doing it, you couldn't.
Just like friends to fill.
It's the right time to actually, it's a
new door to open up to what we can think
about.
Cool. well, maybe we can end with that.
Thanks very much.
Keep your eye on Sam.
Thank you.
Thank you.