Social Network Analysis

Prof. Tanmoy Chakraborty

Chapter - 02
Lecture - 09
Lecture - 04

So far in this chapter on Network Measure, so we have discussed degree you know degree
distribution, we have also looked at different measures which are roughly divided into 3
buckets, right; microscopic, macroscopic and mesoscoping. We have seen metrics like
clustering coefficient, local clustering coefficient, global clustering coefficient, and connected
components, right; strongly connected components, weakly connected components.

So, here we will discuss another very interesting aspect of network measure, it is called
centrality. Now, this is widely used, right.

(Refer Slide Time: 01:03)

Centrality is basically a measure of how central a particular node is with respect to the
network, right. So, you can think of it, I mean if I say that you know to check whether a node
is central to a particular network, so the first thing that comes to our mind is whether the node
is node has equal distance to all the other nodes, right.

Say, if a node has equal distance to all the other nodes, more or less equal distance to all the
other nodes, you can say that you know that node is basically a central node. But we are not

122
talking about that kind of centrality here, we are talking about. Of course, this is one aspect of
centrality, but there are other notions of centralities.

And remember the notion of centrality depends on the particular application. Say, if the
application is say outlier detection, you have different notion of centrality. If the application
is information spreading, you have another type of centrality. If the application is you know
vaccinate nodes, you have another type of centrality and so on and so forth.

So, in general when we talk about centrality I generally look at it you know using this 3 these
4 Ps, right. So, the first P is the prestige. We want to quantify prestige in the formulation of
centrality. We will try to quantify prominence in the formulation of centrality. We quantify
importance of a node in the formulation of centrality and then power, ok. So, these 4 Ps are
something that we will try to incorporate in the definition of a centrality, ok.

(Refer Slide Time: 02:46)

So, let us look at the first centrality, the basic one, the first centrality measure which is called
the degree centrality. So, we know what is degree? Degree is basically the number of edges
adjacent, number of edges which are incident on a particular node, right.

So, what is degree centrality? Degree centrality is essentially the degree of a node, but it is
normalized, right. So, degree centrality of a node v is the degree of the node v and divided by
the maximum degree of a node present in the graph, right. You see the max degree and that is
the denominator.

123
So, for example, in this particular graph G 1, if you look at say node 1, the degree centrality
of node 1 would be 2 by the max degree is 3, right. So, the max degree is 3 or 4, right, the
max degree is 4. So, this would be 2 by 4, ok. Degree centrality of node 3 is 4 by 4 and so on
and so forth.

So, as you can understand it basically ranges between 0 to 1, and the more the degree
centrality the higher the case that, I mean the higher the likelihood that the node has the node
has maximum degree. And therefore, you know in certain applications for example, say you
want to identify prominent node or you want to identify celebrities for example, right.

And so, I mean in any application where you use degree as a measure, right, you can use
degree centrality. That is a very simple notion of a centrality.

(Refer Slide Time: 04:33)

The second one is called closeness centrality. Remember, all these central centrality measures
that we are talking about here, these are all node centric properties, but of course, you can
map it to its edge centric as well, right.

So, the next one is closeness centrality. So, as the name suggest we will look at how close a
particular node is with respect to other nodes in the graph and depending on that we define
this closeness centrality.

So, closeness centrality of a node v is basically, I mean you can basically say that let us look
at the shortest path distance, right of say distance of v from the remaining nodes u, where u

124
belongs to the vortex set V minus the v itself, right because we will not we will not measure
the distance of v with respect to itself.

So, with all the other nodes, what is the distance of v and you know you just normalize it in
some ways, right. If you do it, that is also ok. But the problem is that we also want that higher
the value of centrality better would be the node, right. Higher the value of closeness centrality
better would be the quality of the particular node.

Say, think of a node, right which is closer to all the other nodes. So, the sum of the distance
would be lower, compared to another node which is farther you know which is farther from
the other nodes, right. So, say let us say you have u and v, u is closer to all the remaining
nodes, so the distance would be the sum of the distance would be lower whereas, v is farther
from the other nodes, so the distance would be higher. Therefore, the closeness centrality
value, forget about the denominator, you can use any denominator, right. I will discuss.

So, irrespective of the denominator the closeness centrality value would be higher for node v,
because for node v the numerator value is higher, right. But that is not the desired output,
right. We basically want that node; the node which is closer to the other nodes should have
higher centrality values, right.

So, how can we then you know make to do some sort of tweaks here in this particular
formula? We actually reverse it, ok. So, closeness centrality is basically defined by the
fraction where the denominator would be the sum of distance, ok and the numerator would be
the normalizing factor.

So, in this case the normalizing factor is number of nodes minus 1, mod V minus 1. So, if you
do this, in that case, lower the distance lower the sum of distance higher the centrality value.
So, the node which is at the center at the center position of a network, which has lower
distance with the other with the remaining nodes that will have higher closeness centrality
value, ok. So, using the closeness centrality, we can essentially identify nodes which are
central to a particular network right.

What is the uses of this particular centrality measure? Now, think about it. So, I am basically
trying to identify a node which has small distance with the other nodes, right.

125
So, it means that you know if I want to spread, say, for example, I want to spread up
information a particular information to the entire network, ok, I would basically identify a
node, right which has higher closeness centrality, right. If I identify that node and if I
convince that node to spread the information, it is highly likely that the information will
spread across the network faster than the case where the node has lower closeness centrality
values, ok.

So, in case of say viral marketing or say spread of information, so spread of fake news
misinformation, right in fact, spread of say any sort of epidemics, right. So, you basically I
mean from an adversarial point of view, think of it as an attacker you are an attacker, and you
want to attack the network by spreading the fake news piece information, you basically
choose nodes whose central whose closeness centrality value is maximum, and you basically
convince that nodes to spread misinformation, ok.

(Refer Slide Time: 09:35)

So, now let us move to the; so, I have already you know discussed I mean how to formulate
closeness centrality. Let us take an example. So, let us say we have G 1 as a graph, right and
let us look at the closeness centrality of node 1. So, node 1 if you look at the distance from 1
to 2 the distance is 1, from 1 to 3 the distance is 1, from remember we are interested in
shortest path distance, ok. We discussed what is shortest path distance in the last lecture.

126
From 1 to 4 we have distance to this and this, right and from 1 to 5 we have distance 2, right.
So, the sum of the distance would be 6. Total number of nodes is 5. So, 5 minus 4 is 5 minus
1 would be 4, so 4 by 6 you have 0.67, right.

But what about 3? So, 3 has distance one to all the other nodes from 1, from 2, from 4, from
5, right. So, you see the denominator is 4 and the numerator is also 4, therefore, the closeness
centrality of node 3 is 1, right. So, you can also look at it manually that 3 is basically at the
center of the network, right, therefore, its closeness centrality is maximum.

So, if you want to spread certain information you basically need to convince node 3, and if
node 3 you know spreads some information or to its something, right immediately it is one of
neighbours will receive that information, ok.

(Refer Slide Time: 11:15)

Now, let us look at this third notion of centrality which is betweenness centrality. So, this is
very interesting. So, as the name suggest we will look at nodes for a particular node, we will
basically look at you know possible paths and see whether this that particular node is a part of
the paths, right.

So, let us try to quantify it. So, what we will do, we will take all pairs of nodes. There are
in-situ pairs we take all pairs of nodes and for every pair for every pair x, y for example, right
for every pair x, y we look at the shortest path. Now, remember there can be multiple shortest

127
paths, right the value of the shortest path would be unique of course, but you will have you
can have multiple shortest paths.

So, for every pair x, y I will figure out the shortest paths and I will see whether the particular
node v for which I am measuring the between a centrality, whether node v is present in how
many of the shortest paths, right. Say, between x and y there are 5 shortest paths, and out of
these 5 shortest paths, there are 3 shortest paths where the node v exists. These 3 shortest
paths actually move through v, right and that would give you the quantity.

Now, we will do this for all pairs of shortest paths, right. So, let us assume that the sigma xy v
is the; so, let us first look at sigma xy. So, sigma xy is the number of shortest paths between x
and y, number of shortest paths between x and y. And I am interested in computing the
betweenness centrality for node v. So, out of all the shortest paths between x and y, right how
many shortest paths, in how many shortest paths this node v exists?

So, this sigma v xy indicates that out of all the shortest paths between x and y, how many
shortest in the how many shortest paths between x and y the node v exists, ok. And I basically
do this for all pairs, x comma y which is a part of V cross V for all pairs, ok.

Now, if now think about it. If a node has higher betweenness centrality, what does it mean? It
means that the node is present in almost all the shortest paths between pairs in the network,
right. So, whatever path you want to choose that particular node will always come, that
particular node will always you know be encountered. What does it mean? And where this
betweenness centrality would be useful?

So, think of, you know think of a vaccination drive, right. So, say an epidemic is going on
and you want to you want to stop the spread of epidemic, so and you want to vaccinate
certain nodes, ok. Now, think of a think of a network like this, right like this, ok.

Now, think of this node, and let us assume that the epidemic has been detected in this part of
the network, in this part of the network, right, and you and the epidemic has already been
detected. And you want that the epidemic will should not flow from this part of the network
to this part of the network.

128
So, what would you do? You basically vaccinate this node because you know that this is the
junction point, and if you vaccinate this node, right it will the epidemic will not move from
this part of the network to this part of the network, ok.

So, if you compute the betweenness centrality of all the nodes in this particular graph, you
will see that the betweenness centrality is maximum for this junction node, which is very
obvious because when you take all pairs, right when you take one node from this component
and one node from this component, you will have to move through this particular node, right.

Now, this kind of node is also useful for identifying clusters, ok. Say for example, if you
identify nodes which has higher betweenness centrality or the highest betweenness centrality,
and if you remove that node, what would happen? You may see that the network would get
disconnected.

Say, for example, in this case if you remove this node the adjacent edges will also be
removed, right. So, you will see there are two you know disconnected components emerging.
So, this betweenness centrality concept is also useful for the purpose of attacking a network.

Say, you want to decentralize a network, now think of a terrorist network. So, let us think of a
terrorist network, where nodes are terrorist and links are relations between terrorists and you
identify a node which basically acts a bridge between two terrorist organizations, right.

If you break the node, if you say if you kill the terrorist or if you arrest the terrorist, what
would happen? The information that is flowing from one group to another group that will be
broken, that will be stopped, right. So, these are some of the some of the applications of
betweenness centrality.

Now, this node whose betweenness centrality is maximum, this is also called articulation
point. Now, this terminology is very very useful, right. So, articulation point is a node, whose
removal will disconnect a particular network, ok. It is not necessary that if you remove a node
it would disconnect the network, right. But let us say you have another edge like this, ok, so if
you disconnect this node nothing would happen, the network will still remain connected.

But at least what would happen is that your cluster would become prominent, ok meaning
that your inter cluster edges, so these are inter cluster edges, the inter cluster edges will be

129
reduced and intra cluster edges will remains same. Therefore, clusters the clustering structure
will become prominent, ok.

So, and this is also useful for and this you know articulation points or the nodes with higher
betweenness centrality. These nodes are also very dangerous. Why? Because say for example,
you want to you want that some of the information should not you know should not move
from one component to another component, right.

So, and this these nodes would act as a spy, ok because this is the only node which connects
two different components in a network. So, you should be very careful you know from this
nodes. You should not you know, you may not want to share critical information with this
node. These are also called gatekeepers. Gatekeepers, articulation points, these are these
terminologies are used in you know different applications, right.

(Refer Slide Time: 19:07)

So, let us take an example. So, say this is G 1, and what we are doing here we are creating a
matrix, this is a matrix 5 cross 5 matrix, you have 5 nodes here 5 nodes here, right. And a cell
say x bar y, this is basically used to indicate that say.

So, what is y? y is the let us say let us say let us look at this component, ok, 0 slash 1. Means
that there are two shortest paths, between the there are one shortest path, between the, I mean
there is one shortest path between 1 and 2, and there is no shortest path between 1 and 2
where the particular node v exists.

130
So, I want to compute the betweenness centrality for node v which is 3, in this case. There is
no shortest path between 1 and 2 where the node 3 exists. Let us look at it. So, between 1 and
2, you see that there is only one shortest path and 3 does not exist, right.

Let us look at this one, say between 2 and 4, between 2 and 4 you have this shortest path and
you have this shortest path, ok. So, among these two shortest paths. So, therefore, this is 2,
then the denominator is 2, and among these two shortest paths in this shortest path which is 2,
3, 4 the node 3 exists. So, therefore, the numerator is 1, right. So, the fraction is 1 by 2, right.

Similarly, we do the calculation for all pairs between 1 and 1, 1 and 2, 1 and 3 and so on and
so forth. And then what we do? We basically take the sum. So, in this case, 0 by 1 plus 0 by 1
plus 1 by 1 plus 1 by 2 plus 0 by 1, 0 by 1, sum of all the cells. And for node 3, it would be 4,
ok. So, this is betweenness centrality.

So, sometimes we mix up the concept of betweenness centrality and closeness centrality.
Think about it very carefully. These two metrics capture two different notions.

Closeness centrality captures how close you are with respect to other nodes whereas,
betweenness centrality indicates you know what is the what is the likelihood that whenever I
basically want to move from one node to another node you will be encountered, ok. So, you
cannot avoid the particular node in every shortest path distance calculation, ok. There are two
different notions. I hope you understand what I am trying to say, alright.

(Refer Slide Time: 22:01)

131
So, now this betweenness centrality can also be computed for edges, right. So far we have
computed betweenness centrality for a node and you can easily extend it for edges because
straight forward, right. So, now, again you look at all shortest paths, all pairs of shortest paths
and see for a particular edge, how many times that edge exists, right in all pairs of shortest
paths and the calculation will remains same.

There is another notion called flow betweenness centrality, right. So, in betweenness
centrality, we are only looking at the shortest path shortest paths between two nodes.
Sometimes you may not be able to move through shortest paths because it may happen that
shortage paths are you know are congested in some ways, you may not be able to use the
shortest paths. You may want to use some other paths which may not be shortest path, but
you know slightly longer, but that would still you know fulfil your purpose.

So, in flow between a centrality instead of looking at the shortest path, I would actually look
at all possible paths between a node, right between a pair of nodes. So, all possible paths
between a pair of nodes is computationally expensive. Therefore, we generally do not
calculate flow betweenness centrality, but for certain applications, right we may want to
calculate flow betweenness centrality, ok.

(Refer Slide Time: 23:35)

Now, we are moving to a bit complicated centrality measure. So far we have looked at you
know simple measures like distance, shortest path distance and so on and so forth degree.
Now, we are moving to the to another notion which is called eigenvector centrality.

132
And believe me or not this eigenvector centrality is something based on which a series of
centrality measures have been proposed, ok. So, what is eigenvector centrality? So,
eigenvector centrality says that you know when we measure the centrality of a node, you
would also look at the other nodes which are linked to the given node.

Say for example, so this is a network, right you see that node E is linked to many other nodes
in the network, ok. But node B is only linked to, ok node B is also linked to many other
nodes, but node C is only linked to B, ok. And the size of this of every node indicates the
prestige or centrality or eigenvector centrality.

You see that node B has maximum size followed by C, followed by E and so on and so forth,
right. So, what it is saying is that when we measure the centrality of a node, you also look at
other nodes which are actually pointing to this given node and you and you derive the
centrality of the given node based on the centralities of the other nodes.

And the idea is that if I am pointed by some of the highly prestigious nodes, my prestige will
automatically increase, ok. In other words, say on twitter, if I am followed by Shah Rukh
Khan, for example, Amitabh Bachchan automatically my follower count will increase, Does
not matter whether the other you know other users are following me or not. So, if I am
followed by some you know social media celebrities my prestige will automatically increase,
irrespective of whether others are following me or not.

On the other hand, if you think of another user who is followed by many users, but their
prestige is not that high, my prestige would not be that higher compared to the case where I
am followed by a celebrity for example, ok. And this is the this is a very important notion and
based on that a series of metrics I have been proposed. I will discuss a few in the later slides,
right.

So, essentially you see that you know the same the eigenvector centrality measure is a kind of
a recursive approach, right. So, I am measuring the, I am measuring the centrality of a node
based on the centrality of other nodes. So, I will measure the centrality of the other nodes
based on the centrality of their followers, alright, their neighbours and so on and so forth. So,
it is kind of a recursive definition, ok.

133
(Refer Slide Time: 26:47)

So, now let us move to the definition, right. What it is saying is that a centrality of a node, the
eigenvector centrality of a node v is the sum of the eigenvector centrality of all its
neighbours. So, N v is the neighbours of v, set of neighbours of v and t is one of such
neighbour, so one of such neighbours. So, so x t is the eigenvector centrality of t. So, I am
taking the sum of eigenvector centrality of all the neighbours, right.

And I am actually using a some sort of normalization, say some sort of constant say this is
lambda 1, ok. So, this is the formula. Now, you can write it in a different manner you can say
that, ok 1 by lambda 1, sum of all the nodes present in V, how do I know that the t is a
neighbour of this V, right.

So, I can use the entry in the adjacency matrix. So, a v t, now think of it. So, a is the
adjacency matrix, ok and one such entry is small a v t, small a such sub v t. So, say this is v th
row and t th column. So, this would be either 1 or 0. If v and t are connected then it would be
1, otherwise 0, ok. So, this times x t. If this would be 0, then this will not be considered. If
this would be 1, then x t will be considered, right.

So, now, you can see that I can write it in the matrix form. I can say that this is. So, let me
you know adjust this equation in some other ways. I let me move this lambda 1 here, so it
would be x v lambda 1 equals to sum of, right a v t x t, right.

134
Now, let us make it compact. Let us make it compact; let us. So, x v is the eigenvector
centrality of node v; x t is the eigenvector centrality of node t. Let us assume that I have a I
have a 1D matrix x, where we have elements x 0, x 1, x 2, dot dot dot x v, x t everything.

So, all these entries indicate eigenvector centralities of different nodes. And I have this
adjacency matrix A, right. So, can I write it in this manner? Right. Say lambda 1 X equals to
A X. So, this is basically a and this is X. So, you see this is the equation A X equals to
lambda 1 X.

And if you know matrix operation, right linear algebra you should have heard about
something called eigenvalues and eigenvectors, right. So, this is exactly eigenvalue
eigenvector formula, right equation. So, we know that this is the equation, right A X equals to
lambda X; this is the eigenvalue eigenvector equation, right.

So, if you do not know what is eigenvalue eigenvector, you go back and check, but roughly
what it is saying is that you know eigen; so, let me you know briefly talk about matrix
multiplication. I know this is not a linear algebra course, but let us briefly talk about it, right.

So, A is a matrix and X is a vector, ok. So, now, in a 2D space or a 3D space, right X is
basically an equation of say let us take a 2D, right something like this. So, you can think of X
as a as an equation of a of I mean as a data point here, as a vector here, right a 2D point. So,
when we multiply a vector with a matrix, what actually happens? It basically makes a linear
projection or linear transformation, ok.

So, it. So, when you multiply this by a matrix you get another vector say Y. So, Y is the linear
transformation of X, right and this transformation happens due to A, ok. Now, this linear
transformation. So, what actually this you know matrix indicate. So, if you think of this
columns of this matrix A, right these are basically different you know basis vectors in your
transform space. If you do not know what is basis vector please go back and check.

You basically you are transforming now, in a normal Euclidean space or coordinate geometry
we have two basis vectors i hat, right and j hat, right which is 1 0 and 0 1, these are two basis
vectors, right. Any vector can be generated based on this basis vectors.

135
So, when we multiply X with A, as if A matrix is transforming X to another space, right
where you have two you have two different basis vectors and these basis vectors are basically
the columns of A, ok.

So, during the transformation you will, I mean after the transformation you will get a
transformed vector Y. Now, what is special about this one? This is also transformation A X
equals to lambda X. So, you are multiplying X with A and the transformed vector itself is
basically X, the same vector. So, this is a special type of transformation, ok.

And what is lambda? Lambda is a constant, ok. So, lambda indicates whether you, whether
the transform vector is squeezed or expanded, right. If it is greater than 1 then you basically
expand it. If you if it is a fraction, right, so 0.1 or 0.3 whatever you basically squeeze it. It can
also be negative by the way, negative means you are changing the direction of the resultant
vector, ok. So, this is a special kind of matrix I mean linear transformation where the resultant
vector actually resides on the span of a of the vector itself.

Now, what is span? Again a span, again go back and check the term span, right, basis vector
these are very important concepts, right. So, I am not going into the details of eigenvector,
eigenvalue you know concepts. But what it basically does in this particular case you see that
essentially, X which is the eigenvector I mean the eigenvector, right of A, so this actually
captures all the eigenvector centralities, ok.

Now, you may have multiple eigenvectors, eigenvalues. There can be multiple such
transformation. There can be multiple such vectors which you which when you multiply with
A, the vector will remain same with some constants, constant lambda 1, lambda 2. So, which
eigenvector which eigenvector eigenvalue I should consider? So, this is called eigenvector,
this is eigenvalue, right.

So, it turned out that I should consider the principal eigenvector, I should consider the
principal eigenvector. What is principal eigenvector? Principal eigenvector is eigenvector
corresponding to the largest eigenvalue. So, you will have multiple such eigenvalues, lambda
1, lambda 2, lambda 3 dot dot, right. So, you choose the maximum eigenvalue and you take
the corresponding eigenvector, right.

Why so? Your question would be why we suddenly need to consider the principal
eigenvector. Because what I wanted to do is that I wanted to come up with a vector an

136
eigenvector which is whose values are non-negative, right. I want to come up with a
eigenvector whose values are non-negative.

So, how do I guarantee that which eigenvalue I should consider, so that my eigenvector the
all the entries in the eigenvector should be non-negative? Right. There is a very nice you
know theorem called Perron Frobenius theorem, right which basically says that if you want to
identify an eigenvector whose all elements are non-negative, you should choose eigenvalue
which is the maximum, right.

So, therefore, I want to choose the principal eigenvector, right of A. What is A? A is the A is
the adjacency matrix. So, it is a very simple thing. Although, the idea is little bit complicated,
but I mean the formulation is very simple. I have the adjacency matrix. I will get the principal
eigenvector of the adjacency matrix. And each element in the principle eigenvector indicates
the eigenvector centrality of the corresponding node, ok.

So, again if you do not know what is how to calculate eigenvector, eigenvalues, go back and
look at some you know Fundamentals on Matrix Operation, ok.

So, I stop here. The next part of the lecture again that would be a continuation of this chapter.
I will discuss about other variations of eigenvector centrality. I will discuss about PageRank,
card centrality, and so on and so forth.

Thanks.

137