So that would help in that extent.

So depending on the interaction
with the human this talk.

I will cut you some tough tough clear
mining and A for each type of pasta.

We will see if you stop
Tusker with bacteria and

hear how this possibly attributes

according to different types of jobs that
he grasps but it's time evolving time.

Let me go out and
holler the map to my cities to work.

So in the rest of this present.

I will go through each other's cell one
by willing to tell the first year for

the past aquarium.

Before tabbing to the particular
focus on creating

I want to mention that the main two for
creating logic grabs you proximity human.

So in the next few farmers I'll help
you go through some bag wrong about

the proximity of men over though I know
you might miss to tell you can mistake

to talk up out of the bag wrong
such that in time you talk anyway.

So given the job.

We want you quantify how close or
how rather than we can

do you know without these graphs
without going through to money.

I want you claim then the more question.

You studied good approximate
from another draft.

So here you are in the good sound
hole where we present the proximate

school with each node.

With its back the Quirino
that would afford it.

This go up.

Here we use different the colors to
distinguish different a scored more red

the most relevant to the Quirino and then
we can put all the school with you and

then back to the week with a ranking
back to for the Quirino So

roughly speaking we can think of random
requests to the start as no wine spill.

For example here.

Given the wide clue cause we can spin that
you're off the wine at some location and

then we'll add in this thread
a few too many events from

the city of the many points.

We'll get to some wine.

And so close and they were a civil one and
now four random requests to restart and

we can't think of it as no wind
speed on the class level for

example here on the right side that if
we are interested in the ranking back to

where the fourth Quirino again we can
still be in the Drop a wind as Quirino.

And then we'll let you spread
with a few to the neighborhood.

So eventually that the city of the landing
point the Quirino were received some wind

and so close the neighborhood will
get more wind or rather close.

Most specific to the US suppose that we
have an adrenalin particle just starting

from the Quirino and doing random
surfing on this graph at each step.

I read and will randomly count suites
neighborhood with the probability

that he is proposing though I can wait
between them and also at each step that

when a particle has some small probability
to return to the starting point.

With approximate school for
each node with respect to this clearly no

you define as a steady state of
probability that the random particle will

find that a stay at that
lists here are some simple

animations to illustrate the hollow
when the moon quits or you start work.

Basically it's an event
a move on the graph

where we stand more chance to return
to the creating those at each step.

So in other words they hear the highest
caucus in the school defined by random

to restart meaning that
the we have many Sure and

all highway to the possible
from the source to the target.

OK so enough for
the bag long necks in the first

have tussled queering we want to find
some compact you the specific patterns.

And I would have it you applications.

Here I want to mention
that our algorithm for

both application has to
provide a real system.

I.B.M..

In the first application of Q one the so

called The centerpiece
I grew up discovery.

Give us some clearly notes on the graph.

We want you find a note that have
stronger connections to all or

most of the query notes for
example you may call social class called

the centerpiece guy could be become
a divisor of the querent also rule.

Somebody who had initially
said he searched for

you that to create an awesome be done
to you and assuming that you know you

important on the centerpiece guy
could be the mastermind a criminal or

have stronger connections to all or
most other kind of this suspicious.

Here we have no time to go into
the algorithm that held our centerpiece

the graph discovery that one key component
in the algorithm is how would you be mad.

Here the cutest wouldn't even know it with
respect to whole querist that you know

most the street before the case which
we refer to as the end the query

we can measure the proximate
school from each query no and

then we can modify them together
as a good in a school for

the given the so
here you the deed have preceded it.

You come pure defense if you score for
the end the query and

I'm happy to discuss with you.

Health with you.

Line.

And here in the example are here given for
research is that peace and

joy than all of them in that they
hate how mining communities growl and

hand out more they have a side
they have the intro then decide.

So given this for you to hear.

We want to find other research peers who
have stronger connections to walk them.

And here is a result.

Are we all we can see our research is like
Manila the mist and the tragical fight no.

Because of their multiple
disciplinary community.

In the stand application on
Q one we want to incorporate

uses the life for this and I would have
hoped be the back in the quitting process.

And I was sure you weren't
example given also to conference

by of how they grow up we can ask what
has the most of the later conference for.

So here on the left side of the initial A
Yes we can back this ranking this as human

we can tell there are two major groups of
conferences those over the system does

no such labels the radical group is a more
or less a power to a machine learning and

the death machine learning and
AI such as I see out needs.

Each cut.

While the green group is more or
less about that they got bases on such as

figure more the view of the B I C V and
so on.

This ranking there is that you have
come system there with human intuition

because the machine learning and

that they have bases other two major
sub come years his kid you can't rest.

Now you have the user do some magic feat
of X. I.C.M. which she did to win a major

brain going on here in the middle
using your ranking list.

So again if we inspected this ranking
this we can see that all the other

conferences the conferences
information the cure the conferences

have disappeared from the new regulates
Instead it brings down Brown conferences

the conference is information at
your side such as on yank him.

The guy out and so on then you
mention this to also make a stance

because Katie conference also
extracts a log the research is

in view the information that she will
now if they use a fluke give us some

positive feedback on sci guy which
she is a woman major conference

information if you will here otherwise
study you then you rank your list.

So we can see.

That the pause C.I.R.

brings to more wrong conferences
the conference is an informational Q.

aside and here you the details are
preceded with what interactive carrying

the basic idea is that we want
to use the back information

to modify the graph structure so
that we can bias across them and

towards so called about and
also our way from the X.T.

device and again I'm happy to
discuss the details with you.

Offline.

Though that's all for
the first time past our carrying on.

Next he choose we want to do proximity
tracking and here is the motivations.

So far for all of the effort
we have introduced including

the sense of he said discovery into
rocks to carry their posted time to for

the stats he graphs and also we want to
note that the main two for the at pre-K.

Sheens proximity to men in the on the
other hand many of you have to evolve and

over time where you sense we could
have you know it's all address or

that you can see this you know
it's all agast could die out or

the edge awaiting the graphical
changes over time.

So here the question is how can we
uns that you interesting queries

such a time evolving graph for

example here are given also to
conference by how they got over here.

Maybe we want to ask what
are the most of the late conferences

what the hell to get if you or
how close your kid you come to

a believe you conference over here is one
that you solution whether it be to cheer

rack of the proximate score efficiently or
with hide and here is an example

where we're presenting the fit if you use
Paul five conferences up to each year.

So in this way we want
to see how he's made.

Research interests involved over here for

use terms by the year ninety two major
research interest is you can heat up a C.

performance and distribute systems in
the by the year to solve the seven.

Well data bases use you would not use
a major research interest in issuing

an eight how my name also becomes
a stronger research focus forfeit if you.

And here we present the kiddies to
rank with respect to a deal D.B.

conference overhears So in this figure on
the X. axis is the publication year and

the Y. axis is Katie D.C.
rank with respect to field of you.

Conference.
So we can see that with the more and

the more multiple discipline there
are also publishing in post conferences

that they've come mining community and

that they have bases community by getting
closer and closer with each other.

And then in terms of computational
both the ITER to master and

also the street or for the updated
master the US role for online response.

For users are on the phone that
the data is that both master the within

you know more than one hour to put out
put the crack to ranking back to at

each time step to address this issue
are we have proposed in the fast.

Update algorithm which she is based
on two observations from many of you

applications the first observation is
that for many if you don't buy it.

How the graphs.

There are often highly skewed which
means that the number of one types of

object is usually much smaller
than the other type for

example on the D.B.O. P.
data is that we have more than full

Hungary's thousand different
Also the full conference years.

We only have a palace
three point five stars and

differ in the conferences in a similar
way on the food that rated us had.

We almost to have so many million users.

The only eighteen Stalin differ in the
movie is the style of the aging is that

on the in the case that graph evolving
over time less the at each time staff.

We have added to that I
just changes which involved

in one to the different also and
two to that it for in the conference's.

You mean the real applications Thursday of
the update which we define as a minimum

B.Q. and once you die and to Cuba is
usually much smaller then to update you

still for example we may add all of
the new paper from this year's K.D.T.

conference into the V.O.P. dataset and

then the also that you
conference by of how the graph.

We'll have if you Hungary the office
if you saw I was in the active changes

with all of these the edges
correspond to the same conference.

So in this case that blanket of the update
is just one which is a much smaller and

the up they sell.

The problem the faster of the algorithm
used linear with respect to

the rank of the update.

And it has a known quantity lost and

here is its detail the procedure where
the key point is that carefully.

Using the two observations we have made
sure that before we can avoid this or

expensive big M eighty seven
versions in the own life state.

Instead that we can replace them by
some much cheaper operations easier.

By a much smaller meteors in bushing or
by some bases of Ag to

multiplications and that the and
I'm happy to discuss the details with you.

Offline.

So here we come pare the running
time on to real data that.

In this figure the Y. axis that the wall
clock time is in the log is M scale

in our master the head where the body's
much process though that's all for the.

Proxima checking in key story we want
you can Pugh the proximity quickly.

So here the form of affirmation for
when the Marquise to restart where I was

a ranking back to where the I squaring the
W. denormalized adjacent some eighty S.

of the graph one minus see
is a restart of property and

the starting back to for
the ice Quirino So you read them all.

Chris restart we always have one element
in the starting back to be a one and

a zero is for all the US where
you are here if we are interested

in the ranking backed up with
a force queried know what we have

stairs are forced element in
the starting back to you.

One in the arrows for all the others.

And then the question is
given the Quirino the I or

Ukraine to waste a given a starting
back to I how can we solve

this in your system to make
a long story short her.

It turns out that in order to
come Putin the more queer story

starts with a really mild to use in
this or quote a cue made to yes which is

the worst of that of
an identity mateys minus C.W.

it turns out that the corresponding
element in this Q. maze he is is

a propulsion ode to the proximate school
with between two nodes in the graph for

example here the the first column of
this Q. made here is the propulsion

though to the ranking back to where
the four square in the in the crowd.

And then the question is how
can we get these Q. made here.

So at least to the element of this
cumulative they have to struggle for

the master of both of them
have something brought for

example of the Pauline mass so that maybe
still all four on Linus Pauling things we

may need a lot of help mate is a factor
modifications to get comfort us

on the other hand are the three
come here to master the art.

Prohibitive of flying cost in both time
and space and then the half in years.

How can we balance between the two sides.

The offline cause was
his online course and

our goal is to be fishing to get
the element this bigamy to sing worship

the purpose of mass so
that is quite appealing algorithm.

Here is the basic I.Q. in building our
first we try to find the communities and

the craft and then we'll try to come pure
this door within communities course.

And then in the own life stage.

First we try to find the community
that clearly know the prolonged to and

then resolve those within
communities course and

then they will try to fix
the influence of those bridges in

a fine then they will come by the two
parts together to put the final ranking

back to the most specific of the key
operations in the algorithm of this so

called a proc mean of the accomplishing
on the normalized adjacent summated W..

That he is we want to approximate that W.
by eight Brock diagonal mateys W. one.

Here the prudent way to which correspond

to the within commuting class
is already approximation.

Here the green way to see is.

Which correspond to use always cross
community links so by such walking

near you come to showing the maze he said
we want to universe has this formula.

So here we want to know that
the roommates he uses it to be

inferred to the things it's a product
I don't know mateys we can't do

the inversion independently for
each of them has no profit in the in

addition to the green parts per
the Beijing to the roommate.

So based on the store color
Sherman morning we can get

the inverse of the whole equation.

That the die.

The queue maze here based on
the worth of the roommates he's.

So to summarize billing in
the precompute they will try

to prick on Putin this a few but
a much smaller Beatty's in potions

and then in the online
stage we only need a few

rather than many mateys of Ag to odd
occasions your fine is a ranking back to.

And here we present on the on line clearly
a time versus precompute patient time

in log log scale and
ideally we want the algorithm in the left.

The bottom corner.

So here are our results which is a model
by that weather so cold so rough.

Really.
Hence the sheaves need good a balance

between the online course
because it's computational cost.

So you building on the basic
idea is to efficiently.

So the speed up with respect
to the two St are for

the masses that lie here on the fly
method basically is appalling masses and

start year precomputed

young we come to that Suppose
that we can compute the end.

The priest or the whole mateys THEN YOU'LL
BE own lies that you are just that you

read out across one color right.

So it's unlikely but maybe the not

good good.

So you're building the basic idea
is that you fission approximate

when larger than your system and the a lot
of other interesting new social questions

incomes are proximity come here for
example if

our craft is in a skew to five
hundred in other words if we in

choosing a compact the other media system
is allowed to use a credit with small.

Or how can we leverage this fact that
you fish in the computation and also

in stance to not use we may interested
in a lot of proximate cause for example

maybe we want to use approximate here as
the input of some cost in the algorithm.

So there the trying to use how can we.

Some pain yesterday solve multiple but

the whole four into a call eight
having a system efficiently.

So here we'll have no time to go into
the details of each of this algorithm for

the part when I knew that building user
pays off all these fairy tale trees.

So that's all for the POS of queering
next to the footpaths of mine

in the first minute we want to
immunize a shooting given the graph.

You don't associate that whole
corner come in that hold and

also the virus provocation motto.

We want to stay lactate Bassam know it for
United States you so

that we can prevent and have them from
breaking out to the maximum extent.

We focused on the if we like
a lot of the publication model

the so-called as the highest model.

It turns out to for
this particular virus of publication.

What is really bad to use a D.D. value
of the Education Committee of the God

that he is given the stress of
the eyes attack illogic that he didn't

even by a do you mean that the graph is
the more vulnerable to the virus attack.

It's the more likely
an epidemic will break out.

So based on this observation and optimal
immunization strategy to sit lactic

a node whose absence at the leashing
will create a larger secure job

in terms of the leading agan body paella
this is a massive That is a computational

intractable because of this continent for
example you are on the medium size.

Graph with only one style of in though

we were the new number then two thousand
years to find the best five notes for

immunization on to address
these issues we have proposed

an algorithm called a national and
there are two main ideas behind national.

First of all every time with the veto from
the aren't you know ground when you look

it up can be few that as in a perfectly
to version of the origin of life and

therefore the new leaving a good
bye to you can be approximated

by the agan by you and
A going back to the hour.

And you know graphs the intrusion of such
approximation is that we want to find

a subset of K. nodes which are one
size each of them has high scores and

on the other side and they have this
similar or disconnect among themselves.

And then the approximate age is still
called the seven molecule which

means that we can develop something that
you think if we could leave you with

respect the size of the year we get
some new optimal result for example.

Again on the same graph
with the one thousand and

note within that I sure would.

We only need a less than zero
point zero one stagnant.

If you find the best five
in the United States you.

Mesure that is a new optimal which

means that our solution is that within
the fixed a fraction from the optimal way.

This is due to the Story Corps
the thermometer property and

he even a quick reminder.

So are these graphs suppose out that
we have already had to Reno and

we wanted to get her to ruin those
on the five and six the buy some

molecule is the marginal benefit of these
using these rules no you creases with

the size of the green stat the subset how
know that we have already seen you tip.

And that the function we want to
maximize is something modular then

a case that agreed the algorithm
can guarantee a one minus one or

where he new album or solution.

And here are let me also give
you no clue to prove why

the approximate age used up March or
so again on this graph.

I suppose that we have
already as you stumped Reno.

Now we want to delete the two
more notes on the order five and

six the marginal benefit of
deleting all the five and

No six can be written at this
formula which basically has

two terms the brought home is the pew
benefit of the Legion of the five and

six as the person with whom is
the interaction between the green stat and

the proof that they hear both
terms are not beggars here and

here we are wrong to notice
that only the purple terms.

Depends on the greens that the stuff says
he knows we have already defeated and

therefore if we have already beat you to
a larger green start such as the on C.

drive graph.

Then we'll have more interactions
between the green set and the rule set.

And then we will have more personal terms.

Could be substracted from the Pew benefit
of teaching No the five and No six and

therefore its marginal benefit is relative
small or you know other words Margaret.

So here we pass the quality of on
mashie or that in this figure the X.

axis is the number of that they need
to know the number of the vaccines and

the Y. axis large your spatter.

So this figure as a group is the optimal
solution by a compliment to the mutilation

and the rather curve by Nashville is
a barrel close to the optimal solution.

And here we come pare
the difference that you mean I.

Asian policy.

In this figure the X.
axis the time you and the Y. axis this

is the fracturing of the in fact you know
and it's in the lock reserve scales.

So from this figure on the rather
coup by Nash is always the last we

should be in that we always has the least
number of in fact in the in the graph and

in this figure the degree and the equate
hones are the two most of popular

immunization policies in the literature
degree meaning that if we always

immunize Ohio is that if
Reno works as follows.

First of all randomly pick up a node and

then the ruler vendor merely
immunize one of the proper

business sense we are also two so high as
you can really know what you mean eyes.

And here we come pair the running time and
again the X.

axis is the number of the vaccines and
the Y. axis is in the locker

ism scale so that it so that rather
curve by Nash it's much much faster.

And here we pass
the scalability of Nashville.

In this figure the X. axis is a number
of the ads using the graph and the Y.

axis or clock time.

So clearly on that show
the scale of linearly

with respect to the size of the graph.

So that's all for the immunization and
Max in T. five here.

Given that graph we want to
somehow summarize the year or

compress if the find the path things
such as community is anomalies So

on one very powerful tool for
this kind of task is a maze he said.

All right.
Approximation with the edge of the graph

that here is we want to approximate
the occasional tomatoes a by

the multiplication of surreal
rank matrices L. and R..

For example here.

On the left side given also
a conference by how they graph

we can represent here that
added to the made here is

where their goals are the also
the McCollum's that it compresses.

And then we can do a little
rank approximation for

the mateys A where the lab to meet
his Al is often a good indicator for

the fuel costs of membership.

He also group an estimate is that right.

I mean it's often a good indicator for
the call Prince group and the middle meet

his and indicates the interaction between
the ALSO group and the Conference Group.

And on the other hand if we seem high
deviation from such if you're already

a structure which she is indicated
by high reconstruction IRA

is often a good indicator for
anomalies in the graph and

he had a really good sample on it that
all the traffic a data set that they

just had to spend about the turn to
muscle and if the counting is more

than time to solve then the I.Q.
sauciest and destinations and then for

it to traffic in made at each hour we
do measure already approximation and

a process we can trash an arrow
over time in the middle to go.

So here we can see that it is
only construction arrows such as

the used to be indicates the normal
traffic on the other hand the high

street such as the last figure
is often a good indicator.

So there are a lot of

different the mass of the full rank
approximation in the literature.

In our world we mainly
focus on three aspects.

First we want the efficient
the algorithm in both time and space and

then we wanted to be cute too so
that the minute without is very easy

of putting in the final in
the case the graph evolved.

Over time we want to fishing to update the
corresponding you'd already approximate

showing so that we can track a patent
our interest efficiency over time.

It turns out and

none of that you can sitting master
the fully me Sister requirement for

example are always all as to be the
singularity accomplishing is optimal in

both L two and to come pure that you can
is that it has to be it's not efficient.

You're both time in a space and
also things the singular of actors

are the new combinations of
all the are in your columns.

It's not always easy to find
the clue that interpretations for

such columns are for
such singular of actors and

also it's not easy to update a singular
value comes issuing in the case the graph

of evolving over time so
more recently I research is have focused

on the software that you could downplays
a master such as see you are or C.M.D..

Company otherwise as to be D C U.

Our C.M.D. is a much more efficient
in both time then a space and

also sings the use the actual columns and
the roles from the R.N.G.

don't need to use to construct
the resulting the rank approximation.

It's a relatively easy to find
the good the interpretations for

such a column on the other hand there
might use this to break down than C.

in the lab so made to them that I see you
are all C.M.D. conversations which always

supposed time and space and
also it's not a clear they see you.

Our C.N.N. In that case because
he's evolving over time

turf Yes this issues we have
proposed in the family.

You can sample based
approximation cause a creamery.

For this that you graph the bases.

You are to remove that Levy and

redundancy in the latter may
tear down the Buy see you see.

The shootings in the fourth time
you've all been dynamic you have

the basic idea you have to leverage
the submissiveness all the time.

It turns out he has a c across the nation.

I Crissy see you.

Our C N N C's.

Equal to or
better than both see you lot in A C M D.

In both time and
space in a sense of Big-O. notation.

And here is the detail the compass and

for different the mass of from so
he asked back.

Fishing see interpretations and
the dynamics and

hear what you value is the foremost of
Creevy I asked for this that he graph.

So for all these different the mass of
the have the same approximation accuracy

in our master the third rather by is much
more efficient in both time and space.

And here we evaluate the performance
of course D. with time evolving graphs.

So again for

the same approximation accuracy our
mass the much faster than C M D.

Which is the pride best a massive
The fourth time evolving graphs in

the literature.

So that's pretty much
all what I want to stay

in this present a showing on depending
on the interaction with the Hughes.

We have college at two types of
task queering versus mining and

A for each type of past we have seen
if you start the task respectively.

I also do some other work.

So here other top ten.

I personally like some most.

And here use in the country to
overview of my posture walk.

We have seen Stumm of them
in this present Tahsin I'm

happy to discuss it rests with you.

Offline.

So what's next.

As I mentioned at the very
beginning of this presentation.

My general research you
seem to helps a huge.

Do you understand and

that you would logically laugh that a date
hot in there so you doubt my feelings.

Namely quitting mine.

We believe you know possible.

We have made the bet is that
what this long term goal

should I like to come here to
put our own history domination.

Let me give you some examples.

First is the recommendation system

by itself recommendation system
used not in a new research topic.

But the most of that you get this you work
focused on that that are the ones which

mean that you have the queue to the things
that he or she has already know

in my opinion this may not be the bass
that King for the recommendation system.

So instead we like to investigate
this talk with a broader spectrum

recommendation system where they go years
to collaboratively provide a subset of

which can best to capture the horseback
chill of the user interest.

Including their rather than but also maybe
that diversity is not over with hate and

so on and for
a toss up immunization in this talk.

We have focused on one particular
type called by its publication model

the so-called as my as the model.

So what.
How about the other type of bias

propagation model such as I see
this there's that whole thing

factious recovery virus provocation
model such had chicken pox

in an even bigger ten years how
to unionize dynamic setting for

example what if the drug use
all the time the distress

of the by a pack is not big time and

also we want to generalize our own
master to reach the highest on grass.

For example here given
that you recorded it.

Each of you could ask how.

Can we ask that you interesting queries
such as What's the difference between

North America and Asia terms of
different communication patterns for

yourselves and with people he
was an example maybe we can see.

North America teenagers mainly
talking among themselves and

now it's all sort of
hawks among themselves

on the other hand such a kind of
homophily it's not so all of us.

And as a poor posture of
mining again given such need.

Do you coded each of your graph.

How can we find interesting patterns
such as community anomalies

and a fine of it as for the going to be we
like to make a song going to have for us.

I want him.

You want to come here to design
a schedule that is a single machine and

the unit dishing I'd like to investigate
Apollo allows them to handle this

get it free so
let's go back to this the general

research he's seen how to use
healthy hues to understand and

they utilize logical off that eight
eight hundred S. three key components.

Are we sure work on that really we should
always listen to our customers and

we should only keep us get busy.

Mind.

So this isn't my dream to
get you're on to fulfill.

Such a long term research go.

So finally I want to finish
my present by hearing you

one more graph which is my collaboration
that worked during the Ph D. study.

So with that I like thank you.

And the take questions.

Or the words.

The force that that's good client has so
little of the what the human dimensions.

I think one possible way is it's
the privacy issues are that they.

Our STARTING POINT is that we need it.

You Klak those fields.

So link bigger issue.

We are facing is the security
the privacy issues for

example the privacy scene in social
networks but that just when possible.

I don't know that you can.

Yes.

So the question

is what's the relationship between
those fast approximate algorithm

efficient algorithm for
solving that we need a system like

a prick a pre-condition coming to create
a gradient he said but I think that.

Yes the sort

of the so-called Green
function of the eighty's.

I think the major difference is that.

Yes of course there are a lot.

How nice packages for

solving this also mean your system is
passé up with some special genius systems.

But then we found that.

None of them can fully explore to

have the facility importing
the property is real graft some

problems that we are leveraging use that
one use the product I don't know structure

when you see you rank a new causation
structure and if he doesn't use that.

So how is he then that's the.

Secu that distribution
of those Nat who are so.

It's like for most to us.

Those are real graphs.

Only if you know they have so
high degrees.

So after we do those particular things and

a lot of notes we'll I mean we
only have souls within commuting.

So the cost commute
things is almost empty.

So essentially this is what we
are allowed to operate on a much smaller

May two thousand wishing So for example.

At the for the a precondition
coming to the Mideast and

I think it's generally bad hurt that
it's those fixed point matter like that

better than the polling matters
the best deal we need it to do.

A few iterations right.

So is there just a one way that we can
leverage you to come find that you

mess that together that his somehow to
use that precondition the prick hung

cute asian we have did that as
sort of the precondition for

the holding a system and
then we started to do.

It's also a complicated
green you stand and

also instead of solving like a wing
particularly in your system we also have

the algorithm to solve life multipole
the whole for you to call it even your

system which I didn't go through and
need help that we can discuss the details.

If you want to work.

Yes I have.

Yes I suppose you specially after
those potted Sheens the cost

community just as a cross community
made the very best and a lot of them.

I just empty.

Which means that the ONLY HAVE THOSE which
is communicating this I agree with you.

They have some possible ways we tend
to leverage it to some how to design

some huge mistake.

Maybe to make that for the increment.

To invert the ME.

Yeah so yes right.

This Pentagon massively that you
that you met that I'm not sure

what the particular massive thought here
we cannot avoid that May two thousand and

Bush in operation only that we can
do it on the much just eighteen.

So caught up when the P five
eighty if they just do the same

thing we need that you do in the May two
thousand version will what you diagonal

drop what you community lot that we need
an additional May to university as well

as old guard committed things that all of
them hold more to compare with the orange

you know side of God And

so you mean like even after popping to

be logged on agonal it diagonal
rock if you pants are right.

So I think about it's one possible way
to speed up this proof that you that

we can't give out of a rock equal
great thing which is that you notice.

The wrong kind of India we can apply
it to we could be upright that

our master again but that you thankfully
no matter you should provide and

prod him the way he would like
you may teach you in your case to

be each come is the logic
that you want to get

bookings we can apply to
feel how GREAT them could

be that your venture will have life
it hydride cold like an algorithm.

Yeah that's right.

See how do we interpret the middle mateys
about it because it usually is this one is

compact there and.

And the dance and
also an unnecessary non-negative life.

So can we do not have a better
solution to create the middle may

to use the the seniors out there.

The laughter may taste and
the dried meat is something that we

have to call him from now until May
two thousand wrongly in the future to.

The Cretaceous of the middle may tease you
to see which not nag me to affect you for.

After we cause those samples of growth and
a call him to what's the all to mold

middle matey is giving the comfort
that we want to make sure that

the elements are dismayed to see it's not
an act you maybe we can solve is taking

some like the use of Gresham's away
with Nandini and aggression the way.

Yet for our particular application
like a normal detections it's

not a big issue because after we fungus or
anomaly we see mainly we used to.

So this is sung political with a sample to
call them to interpret to have resolved

but I think maybe in some more contacts
the application with voice sample for

community it happens then
the interpretation for

the middle mates he's more important
because it's a really indicates

how you all will cost in the column
cost interact with each other

then we'll have difficulty like what do
you mean by that not even active that.

So the question you that
if we can afford it.

You sound cold the whole side of the
columns and there you go with the whether

or not we can recover the aren't you know
matey is that and then the answer you.

Yes because for all these examples
based on math as the face the idea that

we want to sample some I would sure
hormones from the arms of graph

which is usually a subset
of the illegal column and

then we somehow we want to project the
orange in our mates into the server space.

It's a band of by those sample the column.

So in this situation you
imagine if we have that column

then if we project at the on your mates

he is that it's getting his that we aren't
YOU made the counting that in the face.

I use to help you you know
we can discuss and we

were.

You were

there.
You were

OK

So two questions on the first
one you that whether or

not we can give some theoretical
analysis all get on with it.

We have developed.

I think generally it will be
difficult because instead of

we we delegate and pre-code evaluations.

If we have the gong to choose what they
have how then we can do it so long have

you valuations Otherwise we do some case
studies before I think passed up immunize

a shing the answer is yes we can get this
woman getting keys and the will not.

It will be to work at the four and
piped up grabs.

So and out of three.

So it turns out that we want to
minimize that meeting a value and

then if you want to approximate the way
and we have a woman to get here and

the good data to us then of
your lightning social network.

If we want to sail just into the field you
see how do we actually come back a week.

So what's the graph that on there
which we are to the United Nations.

So we really do not have
good solutions for that but

then we have something
an issue contact with you.

The monitors that you see and
you there was a way we are trying to solve

this how it's used on the nation with that
so they have this nice to the so-called

age in the face a model to think in
a to give it heat of the whole thing.

So I hope we're going to get help in
the morning and I went to school and

then had a few social.

You bet a few days and

maybe also then after school he goes
back to playing in the playground.

So after we're done.

It is to me because that if you
come to shock the eath reasonable

social network based on that time we
do not have this to do with you for

getting such kind of grabs.

OK So

the question is the what
is the thought of them.

So the stock

market is also all known as
the diminishing return you mean

that it function if it's
the monotonic crease.

But has the on the to be able to out this
month Pike increasing sort help decrease

is that once you have more and the more
upset that I had already the present

quote you from somebody you can reach and
then you.

Give the human more money is of
course that you know you are cruel to

live quite up to him but I think you
would read his more of the compact

the if you give the theme among health
money to some pork you cold like me.

So the politician have this to soak
with a moderate molecule lad he

property then use a knife theory and
I think it's almost like a study

that you can use some case that agreed
the algorithm and the optimal results for

the case step agreed to works at followers
but it's not good for the algorithm.

So every time we try to do it in
order to unite and if we want to

know that we have already done that at
every time we want to find the note which

has the marginal benefit to what you
know ideation so on cue we find it.

OK No At the end then we have done so
the nine seven eight results.

Even when they can be the Fed

in the United Nations the after
we do that approximation.

You can also have the it's nice.

So the theory of the thermometer
that it he is not our contribution

not quality so that the follow
it up from the ninety seven a

no no that's good that the algorithm
is a general for any type of

graph and we do not have those countering
that only it will work and hyped up crap

so we sure that this will
have a nice a mug of hot.

Yes

yes correct.

Yes I think so the question about what's
the difference between a random walk with

with some other possible variants
like a case that her truncated or

in the more crispy stuff.

So we could have if you actually
could have a few other possible

variants such as like.

The saw quick escape of the older
man with some universe or things.

So our experience is that more or less
the all of these different imagine him and

will give you more or less the same
performance in the sense that the or

try to summarize the multiple races that

cannot be seen to know that you can sample
our friend the more Chris we started and

as a truncated random walk with the
thought that you remember that it turns

out that the child to summarize all
possible path from store to height and

then you the chunk of pushing it's Ukraine
that you see that we do not want to come

through the Suez Canal on par with
the last year because then Kate so

another nice thing about this year that
I will see all these different variants

under a date or two or similar to random
or crease restartable by related to

some measurement it can become purely
based on random or three stars or

pseudo random or quiz restart is that
purely brought up other the area.

And if with some others that we can use
a similar I hear you do not guilty.

Thank you.