C HAPTER 7

M EASURES AND METRICS

An introduction to some standard measures and metrics
for quantifying network structure, many of which were
introduced first in the study of social networks, although
they are now in wide use in many other areas

the structure of a network we can calculate from it a variety of

I F WE KNOW
useful quantities or measures that capture particular features of the net-
work topology. In this chapter we look at some of these measures. Many of the
most important ideas in this area come from the social sciences, from the dis-
cipline of social network analysis, which was developed to aid our understand-
ing of social network data such as those described in Chapter 3, and much
of the language used to describe these ideas reflects their sociological origin.
Nonetheless, the methods described are now widely used in areas outside the
social sciences, including computer science, physics, and biology, and form an
important part of the basic network toolbox.1
In the chapter following this one we will apply some of the measures de-
veloped here to the analysis of network data from a variety of fields and in the
process reveal some intriguing features and patterns that will play an impor-
tant role in later developments.

7.1 D EGREE CENTRALITY

A large volume of research on networks has been devoted to the concept of
centrality. This research addresses the question, “Which are the most important
or central vertices in a network?” There are of course many possible definitions

1
For those interested in traditional social network analysis, introductions can be found in the
books by Scott [293] and by Wasserman and Faust [320].

168
 7.2 | E IGENVECTOR CENTRALITY

of importance, and correspondingly many centrality measures for networks.

In this and the following several sections we describe some of the most widely
used such measures.
Perhaps the simplest centrality measure in a network is just the degree of a
vertex, the number of edges connected to it (see Section 6.9). Degree is some-
times called degree centrality in the social networks literature, to emphasize its
use as a centrality measure. In directed networks, vertices have both an in-
degree and an out-degree, and both may be useful as measures of centrality in
the appropriate circumstances.
Although degree centrality is a simple centrality measure, it can be very
illuminating. In a social network, for instance, it seems reasonable to suppose
that individuals who have connections to many others might have more influ-
ence, more access to information, or more prestige than those who have fewer
connections. A non-social network example is the use of citation counts in
the evaluation of scientific papers. The number of citations a paper receives
from other papers, which is simply its in-degree in the citation network, gives
a crude measure of whether the paper has been influential or not and is widely
used as a metric for judging the impact of scientific research.

7.2 E IGENVECTOR CENTRALITY

A natural extension of the simple degree centrality is eigenvector centrality. We
can think of degree centrality as awarding one “centrality point” for every net-
work neighbor a vertex has. But not all neighbors are equivalent. In many
circumstances a vertex’s importance in a network is increased by having con-
nections to other vertices that are themselves important. This is the concept be-
hind eigenvector centrality. Instead of awarding vertices just one point for each
neighbor, eigenvector centrality gives each vertex a score proportional to the
sum of the scores of its neighbors. Here’s how it works.
Let us make some initial guess about the centrality xi of each vertex i. For
instance, we could start off by setting xi = 1 for all i. Obviously this is not
a useful measure of centrality, but we can use it to calculate a better one xi ,
which we define to be the sum of the centralities of i’s neighbors thus:

xi = ∑ Aij x j , (7.1)

j

where Aij is an element of the adjacency matrix. We can also write this ex-
pression in matrix notation as x = Ax, where x is the vector with elements xi .
Repeating this process to make better estimates, we have after t steps a vector

169
M EASURES AND METRICS

of centralities x(t) given by

x ( t ) = A t x (0). (7.2)

Now let us write x(0) as a linear combination of the eigenvectors vi of the

adjacency matrix thus:
x (0) = ∑ c i v i , (7.3)
i

for some appropriate choice of constants ci . Then

 t
κi
x(t) = A t
∑ ci vi = ∑ ci κit vi = κ1t ∑ ci κ1 vi , (7.4)
i i i

where the κi are the eigenvalues of A, and κ1 is the largest of them. Since
κi /κ1 < 1 for all i = 1, all terms in the sum other than the first decay exponen-
tially as t becomes large, and hence in the limit t → ∞ we get x(t) → c1 κ1t v1 .
In other words, the limiting vector of centralities is simply proportional to the
leading eigenvector of the adjacency matrix. Equivalently we could say that
the centrality x satisfies
Ax = κ1 x. (7.5)
This then is the eigenvector centrality, first proposed by Bonacich [49] in 1987.
As promised the centrality xi of vertex i is proportional to the sum of the cen-
tralities of i’s neighbors:
xi = κ1−1 ∑ Aij x j , (7.6)
j

which gives the eigenvector centrality the nice property that it can be large
either because a vertex has many neighbors or because it has important neigh-
bors (or both). An individual in a social network, for instance, can be impor-
tant, by this measure, because he or she knows lots of people (even though
those people may not be important themselves) or knows a few people in high
places.
Note also that the eigenvector centralities of all vertices are non-negative.
To see this, consider what happens if the initial vector x(0) happens to have
only non-negative elements. Since all elements of the adjacency matrix are also
non-negative, multiplication by A can never introduce any negative elements
to the vector and x(t) in Eq. (7.2) must have all elements non-negative.2

2
Technically, there could be more than one eigenvector with eigenvalue κ1 , only one of which
need have all elements non-negative. It turns out, however, that this cannot happen: the adjacency
matrix has only one eigenvector of eigenvalue κ1 . See footnote 2 on page 346 for a proof.

170
 7.2 | E IGENVECTOR CENTRALITY

Equation (7.5) does not fix the normalization of the eigenvector centrality,

although typically this doesn’t matter because we care only about which ver-
tices have high or low centrality and not about absolute values. If we wish,
however, we can normalize the centralities by, for instance, requiring that they
sum to n (which insures that average centrality stays constant as the network
gets larger).
In theory eigenvector centrality can be calculated for either un-
directed or directed networks. It works best however for the un- B
directed case. In the directed case other complications arise. First
of all, a directed network has an adjacency matrix that is, in gen-
A
eral, asymmetric (see Section 6.4). This means that it has two sets of
eigenvectors, the left eigenvectors and the right eigenvectors, and
hence two leading eigenvectors. So which of the two should we use
to define the centrality? In most cases the correct answer is to use
the right eigenvector. The reason is that centrality in directed net-
works is usually bestowed by other vertices pointing towards you,
rather than by you pointing to others. On the World Wide Web, for
Figure 7.1: A portion of a directed net-
instance, the number and stature of web pages that point to your work. Vertex A in this network has
page can give a reasonable indication of how important or useful only outgoing edges and hence will
your page is. On the other hand, the fact that your page might point have eigenvector centrality zero. Ver-
to other important pages is neither here nor there. Anyone can set tex B has outgoing edges and one in-
up a page that points to a thousand others, but that does not make going edge, but the ingoing one origi-
the page important.3 Similar considerations apply also to citation nates at A, and hence vertex B will also
networks and other directed networks. Thus the correct definition have centrality zero.
of eigenvector centrality for a vertex i in a directed network makes
it proportional to the centralities of the vertices that point to i thus:

xi = κ1−1 ∑ Aij x j , (7.7)

j

which gives Ax = κ1 x in matrix notation, where x is the right leading eigen-

vector.
However, there are still problems with eigenvector centrality on directed
networks. Consider Fig. 7.1. Vertex A in this figure is connected to the rest
of the network, but has only outgoing edges and no incoming ones. Such a
vertex will always have centrality zero because there are no terms in the sum

3
This is not entirely true, as we will see in Section 7.5. Web pages that point to many others are
often directories of one sort or another and can be useful as starting points for web surfing. This is
a different kind of importance, however, from that highlighted by the eigenvector centrality and a
different, complementary centrality measure is needed to quantify it.

171
M EASURES AND METRICS

in Eq. (7.7). This might not seem to be a problem: perhaps a vertex that no
one points to should have centrality zero. But then consider vertex B, which
has one ingoing edge, but that edge originates at vertex A, and hence B also
has centrality zero, because the one term in its sum in Eq. (7.7) is zero. Taking
this argument further, we see that a vertex may be pointed to by others that
themselves are pointed to by many more, and so on through many generations,
but if the progression ends up at a vertex or vertices that have in-degree zero,
it is all for nothing—the final value of the centrality will still be zero.
In mathematical terms, only vertices that are in a strongly connected com-
ponent of two or more vertices, or the out-component of such a component,
can have non-zero eigenvector centrality.4 In many cases, however, it is ap-
propriate for vertices with high in-degree to have high centrality even if they
are not in a strongly-connected component or its out-component. Web pages
with many links, for instance, can reasonably be considered important even if
they are not in a strongly connected component. Recall also that acyclic net-
works, such as citation networks, have no strongly connected components of
more than one vertex (see Section 6.11.1), so all vertices will have centrality
zero. Clearly this make the standard eigenvector centrality completely useless
for acyclic networks.
A variation on eigenvector centrality that addresses these problems is the
Katz centrality, which is the subject of the next section.

7.3 K ATZ CENTRALITY

One solution to the issues of the previous section is the following: we sim-
ply give each vertex a small amount of centrality “for free,” regardless of its
position in the network or the centrality of its neighbors. In other words, we
define
xi = α ∑ Aij x j + β, (7.8)
j

where α and β are positive constants. The first term is the normal eigenvector
centrality term in which the centralities of the vertices linking to i are summed,
and the second term is the “free” part, the constant extra term that all vertices
receive. By adding this second term, even vertices with zero in-degree still get
centrality β, and once they have a non-zero centrality, then the vertices they
point to derive some advantage from being pointed to. This means that any
vertex that is pointed to by many others will have a high centrality, although

4
For the left eigenvector it would be the in-component.

172
 7.3 | K ATZ CENTRALITY

those that are pointed to by others with high centrality themselves will still do
better.
In matrix terms, Eq. (7.8) can be written

x = αAx + β1, (7.9)

where 1 is the vector (1, 1, 1 . . .). Rearranging for x, we find that x = β(I −
αA)−1 · 1. As we have said, we normally don’t care about the absolute mag-
nitude of the centrality, only about which vertices have high or low centrality
values, so the overall multiplier β is unimportant. For convenience we usually
set β = 1, giving
x = (I − αA)−1 · 1. (7.10)
This centrality measure was first proposed by Katz in 1953 [169] and we will
refer to it as the Katz centrality.
The Katz centrality differs from ordinary eigenvector centrality in the im-
portant respect of having a free parameter α, which governs the balance be-
tween the eigenvector term and the constant term in Eq. (7.8). If we wish to
make use of the Katz centrality we must first choose a value for this constant.
In doing so it is important to understand that α cannot be arbitrarily large. If
we let α → 0, then only the constant term survives in Eq. (7.8) and all vertices
have the same centrality β (which we have set to 1). As we increase α from
zero the centralities increase and eventually there comes a point at which they
diverge. This happens at the point where (I − αA)−1 diverges in Eq. (7.10),
i.e., when det(I − αA) passes through zero. Rewriting this condition as

det(A − α−1 I) = 0, (7.11)

we see that it is simply the characteristic equation whose roots α−1 are equal to
the eigenvalues of the adjacency matrix.5 As α increases, the determinant first
crosses zero when α−1 = κ1 , the largest eigenvalue of A, or alternatively when
α = 1/κ1 . Thus, we should choose a value of α less than this if we wish the
expression for the centrality to converge.6
Beyond this, however, there is little guidance to be had as to the value that
α should take. Most researchers have employed values close to the maximum
of 1/κ1 , which places the maximum amount of weight on the eigenvector term

5
The eigenvalues being defined by Av = κv, we see that (A − κI)v = 0, which has non-zero
solutions for v only if (A − κI) cannot be inverted, i.e., if det(A − κI) = 0, and hence this equation
gives the eigenvalues κ.
6
Formally one recovers finite values again when one moves past 1/κ1 to higher α, but in prac-
tice these values are meaningless. The method returns good results only for α < 1/κ1 .

173
M EASURES AND METRICS

and the smallest amount on the constant term. This returns a centrality that is
numerically quite close to the ordinary eigenvector centrality, but gives small
non-zero values to vertices that are not in the strongly connected components
or their out-components.
The Katz centrality can be calculated directly from Eq. (7.10) by inverting
the matrix on the right-hand side, but often this isn’t the best way to do it.
Inverting a matrix on a computer takes an amount of time proportional to n3 ,
where n is the number of vertices. This makes direct calculation of the Katz
centrality prohibitively slow for large networks. Networks of more than a
thousand vertices or so present serious problems.
A better approach in many cases is to evaluate the centrality directly from
Eq. (7.8) (or equivalently, Eq. (7.9)). One makes an initial estimate of x—
probably a bad one, such as x = 0—and uses that to calculate a better estimate

x = αAx + β1. (7.12)

Repeating the process many times, x converges to a value close to the correct
centrality. Since A has m non-zero elements, each iteration requires m multi-
plication operations and the total time for the calculation is proportional to rm,
where r is the number of iterations necessary for the calculation to converge.
Unfortunately, r depends on the details of the network and on the choice of α,
so we cannot give a general guide to how many iterations will be necessary.
Instead one must watch the values of xi to observe when they converge to con-
stant values. Nonetheless, for large networks it is almost always worthwhile
to evaluate the centrality this way rather than by inverting the matrix.
We have presented the Katz centrality as a solution to the problems en-
countered with ordinary eigenvector centrality in directed networks. How-
ever, there is no reason in principle why one cannot use Katz centrality in un-
directed networks as well, and there are times when this might be useful. The
idea of adding a constant term to the centrality so that each vertex gets some
weight just by virtue of existing is a natural one. It allows a vertex that has
many neighbors to have high centrality regardless of whether those neighbors
themselves have high centrality, and this could be desirable in some applica-
tions.
A possible extension of the Katz centrality is to consider cases in which the
additive constant term in Eq. (7.8) is not the same for all vertices. One could
define a generalized centrality measure by

xi = α ∑ Aij x j + β i , (7.13)
j

where β i is some intrinsic, non-network contribution to the centrality for each

174
 7.4 | PAGE R ANK

vertex. For example, in a social network the importance of an individual might

depend on non-network factors such as their age or income and if we had in-
formation about these factors we could incorporate it into the values of the β i .
Then the vector x of centralities is given by

x = (I − αA)−1 β, (7.14)

where β is the vector whose elements are the β i . One nice feature of this ap-
proach is that the difficult part of the calculation—the inversion of the matrix—
only has to be done once for a given network and choice of α. For difference
choices of the β i we need not recalculate the inverse, but simply multiply the
inverse into different vectors β.

7.4 PAGE R ANK

The Katz centrality of the previous section has one feature that can be undesir-
able. If a vertex with high Katz centrality points to many others then those
others also get high centrality. A high-centrality vertex pointing to one mil-
lion others gives all one million of them high centrality. One could argue—and
many have—that this is not always appropriate. In many cases it means less
if a vertex is only one among many that are pointed to. The centrality gained
by virtue of receiving an edge from a prestigious vertex is diluted by being
shared with so many others. For instance, the famous Yahoo! web directory
might contain a link to my web page, but it also has links to millions of other
pages. Yahoo! is an important website, and would have high centrality by any
sensible measure, but should I therefore be considered very important by as-
sociation? Most people would say not: the high centrality of Yahoo! will get
diluted and its contribution to the centrality of my page should be small be-
cause my page is only one of millions.
We can allow for this by defining a variation on the Katz centrality in which
the centrality I derive from my network neighbors is proportional to their cen-
trality divided by their out-degree. Then vertices that point to many others pass
only a small amount of centrality on to each of those others, even if their own
centrality is high.
In mathematical terms this centrality is defined by
xj
xi = α ∑ Aij + β. (7.15)
j
kout
j

This gives problems however if there are vertices in the network with out-
degree kout
i = 0. If there are any such vertices then the first term in Eq. (7.15)

175
M EASURES AND METRICS

is indeterminate—it is equal to zero divided by zero (because Aij = 0 for all i).
This problem is easily fixed however. It is clear that vertices with no out-going
edges should contribute zero to the centrality of any other vertex, which we
can contrive by artificially setting kout
i = 1 for all such vertices. (In fact, we
out
could set k i to any non-zero value and the calculation would give the same
answer.)
In matrix terms, Eq. (7.15), is then

x = αAD−1 x + β1, (7.16)

with 1 being again the vector (1, 1, 1, . . .) and D being the diagonal matrix with
elements Dii = max(kout −1 −1 · 1,
i , 1). Rearranging, we find that x = β ( I − αAD )
and thus, as before, β plays the role only of an unimportant overall multiplier
for the centrality. Conventionally we set β = 1, giving

x = (I − αAD−1 )−1 1 = D(D − αA)−1 1. (7.17)

This centrality measure is commonly known as PageRank, which is the trade

name given it by the Google web search corporation, which uses it as a cen-
tral part of their web ranking technology [55]. The aim of the Google web
search engine is to generate lists of useful web pages from a preassembled
index of pages in response to text queries. It does this by first searching the
index for pages matching a given query using relatively simple criteria such
as text matching, and then ranking the answers according to scores based on
Web search is discussed in a combination of ingredients of which PageRank is one. Google returns useful
more detail in Section 19.1. answers to queries not because it is better at finding relevant pages, but be-
cause it is better at deciding what order to present its findings in: its perceived
accuracy arises because the results at the top of the list of answers it returns
are often highly relevant to the query, but it is possible and indeed likely that
many irrelevant answers also appear on the list, lower down.
PageRank works on the Web precisely because having links to your page
from important pages elsewhere is a good indication that your page may be
important too. But the added ingredient of dividing by the out-degrees of
pages insures that pages that simply point to an enormous number of others
do not pass much centrality on to any of them, so that, for instance, network
hubs like Yahoo! do not have a disproportionate influence on the rankings.
As with the Katz centrality, the formula for PageRank, Eq. (7.17), contains
one free parameter α, whose value must be chosen somehow before the algo-
rithm can be used. By analogy with Eq. (7.11) and the argument that follows
it, we can see that the value of α should be less than the inverse of the largest
eigenvalue of AD−1 . For an undirected network this largest eigenvalue turns

176
 7.4 | PAGE R ANK

out to be 1 and the corresponding eigenvector is (k1 , k2 , k3 , . . .), where k i is the

degree of the ith vertex.7 Thus α should be chosen less than 1. For a directed
network, this result does not follow and in general the leading eigenvalue will
be different from 1, although in practical cases it is usually still roughly of or-
der 1.
The Google search engine uses a value of α = 0.85 in its calculations, al-
though it’s not clear that there is any rigorous theory behind this choice. More
likely it is just a shrewd guess based on experimentation to find out what
works well.
As with the Katz centrality we can generalize PageRank to the case where
the additive constant term in Eq. (7.15) is different for different vertices:
xj
xi = α ∑ Aij + βi . (7.18)
j
kout
j

In matrix form this gives a solution for the centrality vector of

x = D(D − αA)−1 β. (7.19)

One could, for instance, use this for ranking web pages, giving β i a value based
perhaps on textual relevance to a search query. Pages that contained the word
or words being searched for more often or in more prominent places could
be given a higher intrinsic centrality than others, thereby pushing them up
the rankings. The author is not aware, however, of any cases in which this
technique has been implemented in practice.
Finally, one can also imagine a version of PageRank that did not have the
additive constant term in it at all:
xj
xi = α ∑ Aij , (7.20)
j
kj

which is similar to the original eigenvector centrality introduced back in Sec-

tion 7.2, but now with the extra division by k j . For an undirected network,
however, this measure is trivial: it is easy to see that it gives simply xi = k i

7
It is easy to confirm that this vector is indeed an eigenvector with eigenvalue 1. That there
is no eigenvalue larger than 1 is less obvious. It follows from a standard result in linear algebra,
the Perron–Frobenius theorem, which states that the largest eigenvalue of a matrix such as AD−1
that has all elements non-negative is unique—there is only one eigenvector with this eigenvalue—
that the eigenvector also has all elements non-negative, and that it is the only eigenvector with all
elements non-negative. Combining these results, it is clear that the eigenvalue 1 above must be
the largest eigenvalue of the matrix AD−1 . For a discussion of the Perron–Frobenius theorem see
Ref. [217] and the two footnotes on page 346 of this book.

177
M EASURES AND METRICS

with constant term without constant term

divide by x = D(D − αA)−1 · 1 x = AD−1 x
out-degree PageRank degree centrality
x = (I − αA)−1 · 1 x = κ1−1 Ax
no division
Katz centrality eigenvector centrality

Table 7.1: Four centrality measures. The four matrix-based centrality measures dis-
cussed in the text are distinguished by whether or not they include an additive constant
term in their definition and whether they are normalized by dividing by the degrees of
neighboring vertices. Note that the diagonal matrix D, which normally has elements
Dii = k i , must be defined slightly differently for PageRank, as Dii = max(1, k i )—see
Eq. (7.15) and the following discussion. Each of the measures can be applied to directed
networks as well as undirected ones, although only three of the four are commonly used
in this way. (The measure that appears in the top right corner of the table is equivalent
to degree centrality in the undirected case but takes more complicated values in the
directed case and is not widely used.)

and therefore is just the same as ordinary degree centrality. For a directed net-
work, on the other hand, it does not reduce to any equivalent simple value
and it might potentially be of use, although it does not seem to have found
use in any prominent application. (It does suffer from the same problem as
the original eigenvector centrality, that it gives non-zero scores only to vertices
that fall in a strongly connected component of two or more vertices or in the
out-component of such a component. All other vertices get a zero score.)
In Table 7.1 we give a summary of the different matrix centrality measures
we have discussed, organized according to their definitions and properties. If
you want to use one of these measures in your own calculations and find the
many alternatives bewildering, eigenvector centrality and PageRank are prob-
ably the two measures to focus on initially. They are the two most commonly
used measures of this type. The Katz centrality has found widespread use in
the past but has been favored less in recent work, while the PageRank mea-
sure without the constant term, Eq. (7.20), is the same as degree centrality for
undirected networks and not in common use for directed ones.

7.5 H UBS AND AUTHORITIES

In the case of directed networks, there is another twist to the centrality mea-
sures introduced in this section. So far we have considered measures that
accord a vertex high centrality if those that point to it have high centrality.

178
 7.5 | H UBS AND AUTHORITIES

However, in some networks it is appropriate also to accord a vertex high cen-

trality if it points to others with high centrality. For instance, in a citation net-
work a paper such as a review article may cite other articles that are author-
itative sources for information on a particular subject. The review itself may
contain relatively little information on the subject, but it tells us where to find
the information, and this on its own makes the review useful. Similarly, there
are many examples of web pages that consist primarily of links to other pages
on a given topic or topics and such a page of links could be very useful even if
it does not itself contain explicit information on the topic in question.
Thus there are really two types of important node in these networks: au-
thorities are nodes that contain useful information on a topic of interest; hubs
are nodes that tell us where the best authorities are to be found. An authority
may also be a hub, and vice versa: review articles often contain useful discus-
sions of the topic at hand as well as citations to other discussions. Clearly hubs
and authorities only exist in directed networks, since in the undirected case
there is no distinction between pointing to a vertex and being pointed to.
One can imagine defining two different types of centrality for directed net-
works, the authority centrality and the hub centrality, which quantify vertices’
prominence in the two roles. This idea was first put forward by Kleinberg [176]
and developed by him into a centrality algorithm called hyperlink-induced topic
search or HITS.
The HITS algorithm gives each vertex i in a network an authority central-
ity xi and a hub centrality yi . The defining characteristic of a vertex with high
authority centrality is that it is pointed to by many hubs, i.e., by many other
vertices with high hub centrality. And the defining characteristic of a vertex
with high hub centrality is that it points to many vertices with high authority
centrality.
Thus an important scientific paper (in the authority sense) would be one
cited in many important reviews (in the hub sense). An important review is
one that cites many important papers. Reviews, however, are not the only
publications that can have high hub centrality. Ordinary papers can have high
hub centrality too if they cite many other important papers, and papers can
have both high authority and high hub centrality. Reviews too may be cited
by other hubs and hence have high authority centrality as well as high hub
centrality.
In Kleinberg’s approach, the authority centrality of a vertex is defined to be
proportional to the sum of the hub centralities of the vertices that point to it:

xi = α ∑ Aij y j , (7.21)
j

179
M EASURES AND METRICS

where α is a constant. Similarly the hub centrality of a vertex is proportional

to the sum of the authority centralities of the vertices it points to:

yi = β ∑ A ji x j , (7.22)
j

with β another constant. Notice that the indices on the matrix element A ji are
swapped around in this second equation: it is the vertices that i points to that
define its hub centrality.
In matrix terms these equations can be written as

x = αAy, y = βA T x, (7.23)

or, combining the two,

AA T x = λx, A T Ay = λy, (7.24)

where λ = (αβ)−1 . Thus the authority and hub centralities are respectively
given by eigenvectors of AA T and A T A with the same eigenvalue. By an ar-
gument similar to the one we used for the standard eigenvector centrality in
Section 7.1 we can show that we should in each case take the eigenvector cor-
responding to the leading eigenvalue.
A crucial condition for this approach to work, is that AA T and A T A have
the same leading eigenvalue λ, otherwise we cannot satisfy both conditions in
Eq. (7.24). It is easily proved, however, that this is the case, and in fact that all
eigenvalues are the same for the two matrices. If AA T x = λx then multiplying
both sides by AT gives
A T A ( A T x ) = λ ( A T x ), (7.25)
and hence A T x is an eigenvector of A T A with the same eigenvalue λ. Compar-
ing with Eq. (7.24) this means that

y = A T x, (7.26)

which gives us a fast way of calculating the hub centralities once we have
the authority ones—there is no need to solve both the eigenvalue equations
in Eq. (7.24) separately.
Note that AA T is precisely the cocitation matrix defined in Section 6.4.1
(Eq. (6.8)) and the authority centrality is thus, roughly speaking, the eigen-
vector centrality for the cocitation network.8 Similarly A T A is the bibliographic

8
This statement is only approximately correct since, as discussed in Section 6.4.1, the cocitation
matrix is not precisely equal to the adjacency matrix of the cocitation network, having non-zero
elements along its diagonal where the adjacency matrix has none.

180
 7.6 | C LOSENESS CENTRALITY

coupling matrix, Eq. (6.11), and hub centrality is the eigenvector centrality for
the bibliographic coupling network.
A nice feature of the hub and authority centralities is that they circum-
vent the problems that ordinary eigenvector centrality has with directed net-
works, that vertices outside of strongly connected components or their out-
components always have centrality zero. In the hubs and authorities approach
vertices not cited by any others have authority centrality zero (which is reason-
able), but they can still have non-zero hub centrality. And the vertices that they
cite can then have non-zero authority centrality by virtue of being cited. This
is perhaps a more elegant solution to the problems of eigenvector centrality
in directed networks than the more ad hoc method of introducing an additive
constant term as we did in Eq. (7.8). We can still introduce such a constant
term into the HITS algorithm if we wish, or employ any of the other variations
considered in previous sections, such as normalizing vertex centralities by the
degrees of the vertices that point to them. Some variations along these lines
are explored in Refs. [52, 256], but we leave the pursuit of such details to the
enthusiastic reader.
The HITS algorithm is an elegant construction that should in theory pro-
vide more information about vertex centrality than the simpler measures of
previous sections, but in practice it has not yet found much application. It is
used as the basis for the web search engines Teoma and Ask.com, and will per-
haps in future find further use, particularly in citation networks, where it holds
clear advantages over other eigenvector measures.

7.6 C LOSENESS CENTRALITY

An entirely different measure of centrality is provided by the closeness central-
ity, which measures the mean distance from a vertex to other vertices. In Sec-
tion 6.10.1 we encountered the concept of the geodesic path, the shortest path
through a network between two vertices. Suppose dij is the length of a geodesic
path from i to j, meaning the number of edges along the path.9 Then the mean
geodesic distance from i to j, averaged over all vertices j in the network, is

1
i =
n ∑ dij . (7.27)
j

9
Recall that geodesic paths need not be unique—vertices can be joined by several shortest
paths of the same length. The length dij however is always well defined, being the length of any
one of these paths.

181
M EASURES AND METRICS

This quantity takes low values for vertices that are separated from others by
only a short geodesic distance on average. Such vertices might have better ac-
cess to information at other vertices or more direct influence on other vertices.
In a social network, for instance, a person with lower mean distance to others
might find that their opinions reach others in the community more quickly
than the opinions of someone with higher mean distance.
In calculating the average distance some authors exclude from the sum
in (7.27) the term for j = i, so that

1
i =
n−1 ∑ dij , (7.28)
j(=i )

which is a reasonable strategy, since a vertex’s influence on itself is usually not

relevant to the working of the network. On the other hand, the distance dii
from i to itself is zero by definition, so this term in fact contributes nothing to
the sum. The only difference the change makes to i is in the leading divisor,
which becomes 1/(n − 1) instead of 1/n, meaning that i changes by a factor
of n/(n − 1). Since this factor is independent of i and since, as we have said,
we usually care only about the relative centralities of different vertices and not
about their absolute values, we can in most cases ignore the difference between
Eqs. (7.27) and (7.28). In this book we use (7.27) because it tends to give slightly
more elegant analytic results.
The mean distance i is not a centrality measure in the same sense as the
others in this chapter, since it gives low values for more central vertices and
high values for less central ones, which is the opposite of our other measures.
In the social networks literature, therefore, researchers commonly calculate the
inverse of i rather than i itself. This inverse is called the closeness centrality Ci :

1 n
Ci = = . (7.29)
i ∑ j dij

Closeness centrality is a very natural measure of centrality and is often used

in social and other network studies. But it has some problems. One issue
is that its values tend to span a rather small dynamic range from largest to
smallest. As discussed in Sections 3.6, 8.2, and 12.7, geodesic distances dij be-
tween vertices in most networks tend to be small, the typical distance increas-
ing only logarithmically with the size of the entire network. This means that
the ratio between the smallest distance, which is 1, and the largest, which is of
order log n, is itself only of order log n, which is small. But the smallest and
largest distances provide lower and upper bounds on the average distance i ,
and hence the range of values of i and similarly of Ci is also small. In a typical

182
 7.6 | C LOSENESS CENTRALITY

network the values of Ci might span a factor of five or less. What this means
in practice is that it is difficult to distinguish between central and less central
vertices using this measure: the values tend to be cramped together with the
differences between adjacent values showing up only when you examine the
trailing digits. This means that even small fluctuations in the structure of the
network can change the order of the values substantially.
For example, it has become popular in recent years to rank film actors ac-
cording to their closeness centrality in the network of who has appeared in
films with who else [323]. Using data from the Internet Movie Database,10 we
find that in the largest component of the network, which includes more than
98% of all actors, the smallest closeness centrality of any actor is 2.4138 for
the actor Christopher Lee,11 while the largest is 8.6681 for an Iranian actress
named Leia Zanganeh. The ratio of the two is just 3.6 and about half a million
other actors lie in between. As we can immediately see, the values must be
very closely spaced. The second best centrality score belongs to actor Donald
Pleasence, who scores 2.4164, just a tenth of a percent less than winner Lee.
Because of the close spacing of values, the leaders under this dubious measure
of superiority change frequently as the small details of the film network shift
when new films are made or old ones added to the database. In an analysis
using an earlier version of the database, Watts and Strogatz [323] proclaimed
Rod Steiger to be the actor with the lowest closeness centrality. Steiger falls in
sixth place in our analysis and it is entirely possible that the rankings will have
changed again by the time you read this. Other centrality measures, including
degree centrality and eigenvector centrality, typically don’t suffer from this
problem because they have a wider dynamic range and the centrality values,
particular those of the leaders, tend to be widely separated.
The closeness centrality has another problem too. If, as discussed in Sec-
tion 6.10.1, we define the geodesic distance between two vertices to be infinite
if the vertices fall in different components of the network, then i is infinite
for all i in any network with more than one component and Ci is zero. There
are two strategies for getting around this. The most common one is simply
to average over only those vertices in the same component as i. Then n in
Eq. (7.29) becomes the number of vertices in the component and the sum is
over only that component. This gives us a finite measure, but one that has its
own problems. In particular, distances tend to be smaller between vertices in
small components, so that vertices in such components get lower values of i

10
www.imdb.com
11
Perhaps most famous for his role as the evil wizard Saruman in the film version of The Lord
of the Rings.

183
M EASURES AND METRICS

and higher closeness centrality than their counterparts in larger components.

This is usually undesirable: in most cases vertices in small components are
considered less well connected than those in larger ones and should therefore
be given lower centrality.
Perhaps a better solution, therefore, is to redefine closeness in terms of the
harmonic mean distance between vertices, i.e., the average of the inverse dis-
tances:
1 1
Ci = ∑
n − 1 j(=i) dij
. (7.30)

(Notice that we are obliged in this case to exclude from the sum the term for
j = i, since dii = 0 which would make this term infinite. This means that the
sum has only n − 1 terms in it, hence the leading factor of 1/(n − 1).)
This definition has a couple of nice properties. First, if dij = ∞ because i and
j are in different components, then the corresponding term in the sum is sim-
ply zero and drops out. Second, the measure naturally gives more weight to
vertices that are close to i than to those far away. Intuitively we might imagine
that the distance to close vertices is what matters in most practical situations—
once a vertex is far away in a network it matters less exactly how far away it
is, and Eq. (7.30) reflects this, having contributions close to zero from all such
vertices.
Despite its desirable qualities, however, Eq. (7.30) is rarely used in practice.
We have seen it employed only occasionally.
An interesting property of entire networks, which is related to the closeness
centrality, is the mean geodesic distance between vertices. In Section 8.2 we
will use measurements of mean distance in networks to study the so-called
“small-world effect.”
For a network with only one component, the mean distance between pairs
of vertices, conventionally denoted just  (now without the subscript), is
1 1
=
n2 ∑ dij = n ∑ i . (7.31)
ij i

In other words  is just the mean of i over all vertices.

For a network with more than one component we run into the same prob-
lems as before, that dij is infinite when i and j are in different components
and hence  is also infinite. The most common way around this problem is to
average only over paths that run between vertices in the same component. Let
{Cm } be the set of components of a network, with m = 1, 2 . . . Then we define

∑m ∑ij∈Cm dij
= , (7.32)
∑m n2m

184
 7.7 | B ETWEENNESS CENTRALITY

where nm is the number of vertices in component Cm . This measure is now

finite for all networks, although it is not now equal to a simple average over
the values of i for each vertex.
An alternative and perhaps better approach would be to use the trick from
Eq. (7.30) and define a harmonic mean distance  according to

1 1 1 1
n ( n − 1) ∑
= = ∑ Ci , (7.33)
 d
i = j ij
n i

or equivalently
n
 = , (7.34)
∑i Ci
where Ci is the harmonic mean closeness of Eq. (7.30). (Note that, as in (7.30),
we exclude from the first sum in (7.33) the terms for i = j, which would be
infinite since dii = 0.)
Equation (7.34) automatically removes any contributions from vertex pairs
for which dij = ∞. Despite its elegance, however, Eq. (7.34), like Eq. (7.30), is
hardly ever used.

7.7 B ETWEENNESS CENTRALITY

A very different concept of centrality is betweenness centrality, which measures
the extent to which a vertex lies on paths between other vertices. The idea of
betweenness is usually attributed to Freeman [128] in 1977, although as Free-
man himself has pointed out [129], it was independently proposed some years
earlier by Anthonisse [19] in an unpublished technical report.
Suppose we have a network with something flowing around it from vertex
to vertex along the edges. For instance, in a social network we might have mes-
sages, news, information, or rumors being passed from one person to another.
In the Internet we have data packets moving around. Let us initially make
the simple assumption that every pair of vertices in the network exchanges a
message with equal probability per unit time (more precisely every pair that
is actually connected by a path) and that messages always take the shortest
(geodesic) path though the network, or one such path, chosen at random, if
there are several. Then let us ask the following question: if we wait a suitably
long time until many messages have passed between each pair of vertices, how
many messages, on average, will have passed through each vertex en route to
their destination? The answer is that, since messages are passing down each
geodesic path at the same rate, the number passing through each vertex is
simply proportional to the number of geodesic paths the vertex lies on. This

185
M EASURES AND METRICS

number of geodesic paths is what we call the betweenness centrality, or just

betweenness for short.
Vertices with high betweenness centrality may have considerable influence
within a network by virtue of their control over information passing between
others. The vertices with highest betweenness in our message-passing scenario
are the ones through which the largest number of messages pass, and if those
vertices get to see the messages in question as they pass, or if they get paid
for passing the messages along, they could derive a lot of power from their
position within the network. The vertices with highest betweenness are also
the ones whose removal from the network will most disrupt communications
between other vertices because they lie on the largest number of paths taken
by messages. In real-world situations, of course, not all vertices exchange com-
munications with the same frequency, and in most cases communications do
not always take the shortest path. Nonetheless, betweenness centrality may
still be an approximate guide to the influence vertices have over the flow of
information between others.
Having seen the basic idea of betweenness centrality, let us make things
more precise. For the sake of simplicity, suppose for the moment that we have
an undirected network in which there is at most one geodesic path between
any pair of vertices. (There may be zero paths if the vertices in question are
in different components.) Consider the set of all geodesic paths in such a net-
work. Then the betweenness centrality of a vertex i is defined to be the number
of those paths that pass through i.
Mathematically, let nist be 1 if vertex i lies on the geodesic path from s to t
and 0 if it does not or if there is no such path (because s and t lie in different
components of the network). Then the betweenness centrality xi is given by

xi = ∑ nist . (7.35)
st

Note that this definition counts separately the geodesic paths in either direc-
tion between each vertex pair. Since these paths are the same on an undirected
network this effectively counts each path twice. One could compensate for this
by dividing xi by 2, and often this is done, but we prefer the definition given
here for a couple of reasons. First, it makes little difference in practice whether
one divides the centrality by 2, since one is usually concerned only with the rel-
ative magnitudes of the centralities and not with their absolute values. Second,
as discussed below, Eq. (7.35) has the advantage that it can be applied unmod-
ified to directed networks, in which the paths in either direction between a
vertex pair can differ.
Note also that Eq. (7.35) includes paths from each vertex to itself. Some

186
 7.7 | B ETWEENNESS CENTRALITY

people prefer to exclude such paths from the definition, so that xi = ∑s=t nist ,
but again the difference is typically not important. Each vertex lies on one
path from itself to itself, so the inclusion of these terms simply increases the
betweenness by 1, but does not change the rankings of the vertices—which
ones have higher or lower betweenness—relative to one another.
There is also a choice to be made about whether the path from s to t should
be considered to pass through the vertices s and t themselves. In the social net-
works literature it is usually assumed that it does not. We prefer the definition
where it does: it seems reasonable to define a vertex to be on a path between
itself and someone else, since normally a vertex has control over information
flowing from itself to other vertices or vice versa. If, however, we exclude the
endpoints of the path as sociologists commonly do, the only effect is to reduce
the number of paths through each vertex by twice the size of the component
to which the vertex belongs. Thus the betweennesses of all vertices within a
single component are just reduced by an additive constant and the ranking of
vertices within the component is again unchanged. (The rankings of vertices
in different components can change relative to one another, but this is rarely an
issue because betweenness centrality is not typically used to compare vertices
in different components, since such vertices are not competing for influence in
the same arena.)
These developments are all for the case in which there is at most one geo-
desic path between each vertex pair. More generally, however, there may be
more than one. The standard extension of betweenness to this case gives each
path a weight equal to the inverse of the number of paths. For instance, if
there are two geodesic paths between a given pair of vertices, each of them
gets weight 12 . Then the betweenness of a vertex is defined to be the sum of the B
weights of all geodesic paths passing through that vertex. C
Note that the geodesic paths between a pair of vertices need not be vertex-
independent, meaning they may pass through some of the same vertices (see A
figure). If two or more paths pass through the same vertex then the between-
ness sum includes contributions from each of them. Thus if there are, say, three Vertices A and B are con-
geodesic paths between a given pair of vertices and two of them pass through nected by two geodesic
a particular vertex, then they contribute 23 to that vertex’s betweenness. paths. Vertex C lies on both
Formally, we can express the betweenness for a general network by redefin- paths.
ing nist to be the number of geodesic paths from s to t that pass through i. And
we define gst to be the total number of geodesic paths from s to t. Then the
betweenness centrality of vertex i is

ni
xi = ∑ gstst , (7.36)
st

187
M EASURES AND METRICS

where we adopt the convention that nist /gst = 0 if both nist and gst are zero. This
definition is equivalent to our message-passing thought experiment above, in
which messages pass between all pairs of vertices in a network at the same
average rate, traveling along shortest paths, and in the case of several shortest
paths between a given pair of vertices they choose at random between those
several paths. Then xi is proportional to the average rate at which traffic passes
though vertex i.
Betweenness centrality can be applied to directed networks as well. In a
directed network the shortest path between two vertices depends, in general,
on the direction you travel in. The shortest path from A to B is different from
the shortest path from B to A. Indeed there may be a path in one direction and
no path at all in the other. Thus it is important in a directed network explicitly
to include the path counts in either direction between each vertex pair. The
definition in Eq. (7.36) already does this and so, as mentioned above, we can
use the same definition without modification for the directed case. This is one
reason why we prefer this definition to other slight variants that are sometimes
used.
Although the generalization of betweenness to directed networks is straight-
forward, however, it is rarely if ever used, so we won’t discuss it further here,
concentrating instead on the much more common undirected case.
Betweenness centrality differs from the other centrality
measures we have considered in being not principally a mea-
sure of how well-connected a vertex is. Instead it measures
A
how much a vertex falls “between” others. Indeed a vertex
Group 1 Group 2
can have quite low degree, be connected to others that have
low degree, even be a long way from others on average, and
still have high betweenness. Consider the situation depicted
in Fig. 7.2. Vertex A lies on a bridge between two groups
Figure 7.2: A low-degree vertex with high be-
within a network. Since any shortest path (or indeed any
tweenness. In this sketch of a network, ver-
path whatsoever) between a vertex in one group and a ver-
tex A lies on a bridge joining two groups of
tex in the other must pass along this bridge, A acquires very
other vertices. All paths between the groups
must pass through A, so it has a high between- high betweenness, even though it is itself on the periphery of
ness even though its degree is low. both groups and in other respects may be not well connected:
probably A would not have particularly impressive values for
eigenvector or closeness centrality, and its degree centrality is
only 2, but nonetheless it might have a lot of influence in the network as a re-
sult of its control over the flow of information between others. Vertices in roles

188
 7.7 | B ETWEENNESS CENTRALITY

like this are sometimes referred to in the sociological literature as brokers.12

Betweenness centrality also has another interesting property: its values are
typically distributed over a wide range. The maximum possible value for the
betweenness of a vertex occurs when the vertex lies on the shortest path be-
tween every other pair of vertices. This occurs for the central vertex in a star
graph, a network composed of a vertex attached to n − 1 others by single edges.
In this situation the central vertex lies on all n2 shortest paths between vertex
pairs except for the n − 1 paths from the peripheral vertices to themselves.
Thus the betweenness centrality of the central vertex is n2 − n + 1. At the
other end of the scale, the smallest possible value of betweenness in a network A star graph.
with a single component is 2n − 1, since at a minimum each vertex lies on ev-
ery path that starts or ends with itself. (There are n − 1 paths from a vertex to
others, n − 1 paths from others to the vertex, and one path from the vertex to
itself, for a total of 2(n − 1) + 1 = 2n − 1.) This situation occurs, for instance,
when a network has a “leaf” attached to it, a vertex connected to the rest of the
network by just a single edge.
Thus the ratio of largest and smallest possible betweenness values is
n2 − n + 1

12 n, (7.37)
2n − 1
where the equality becomes exact in the limit of large n. Thus in theory there
could be a factor of almost 12 n between the largest and smallest betweenness
centralities, which could become very large for large networks. In real net-
works the range is usually considerably smaller than this, but is nonetheless
large and typically increasing with increasing n.
Taking again the example of the network of film actors from the previous
section, the individual with the highest betweenness centrality in the largest
component of the actor network is the great Spanish actor Fernando Rey, most
famous in the English-speaking world for his 1971 starring role next to Gene
Hackman in The French Connection.13 Rey has a betweenness score of 7.47 × 108 ,

12
Much of sociological literature concerns power or “social capital.” It may seem ruthless to
think of individuals exploiting their control over other people’s information to gain the upper
hand on them, but it may also be realistic. At least in situations where there is a significant pay-off
to having such an upper hand (like business relationships, for example), it is reasonable to suppose
that notions of power derived from network structure really do play into people’s manipulations
of the world around them.
13
It is perhaps no coincidence that the highest betweenness belongs to an actor who appeared
in both European and American films, played roles in several different languages, and worked
extensively in both film and television, as well as on stage. Rey was the archetypal “broker,” with
a career that made him a central figure in several different arms of the entertainment business that
otherwise overlap relatively little.

189
M EASURES AND METRICS

while the lowest score of any actor14 in the large component is just 8.91 × 105 .
Thus there is a ratio of almost a thousand between the two limits—a much
larger dynamic range than the ratio of 3.6 we saw in the case of closeness cen-
trality. One consequence of this is that there are very clear winners and losers
in the betweenness centrality competition. The second highest betweenness
in the actor network is that of Christopher Lee (again), with 6.46 × 108 , a 14%
percent difference from winner Fernando Rey. Although betweenness values
may shift a little as new movies are made and new actors added to the net-
work, the changes are typically small compared with these large gaps between
the leaders, so that the ordering at the top of the list changes relatively infre-
quently, giving betweenness centrality results a robustness not shared by those
for closeness centrality.
The values of betweenness calculated here are raw path counts, but it is
sometimes convenient to normalize betweenness in some way. Several of the
standard computer programs for network analysis, such as Pajek and UCINET,
perform such normalizations. One natural choice is to normalize the path
count by dividing by the total number of (ordered) vertex pairs, which is n2 , so
that betweenness becomes the fraction (rather than the number) of paths that
run through a given vertex:15

1 ni
xi =
n2 ∑ gstst . (7.38)
st

With this definition, the values of the betweenness lie strictly between zero and
one.
Some other variations on the betweenness centrality idea are worth men-
tioning. Betweenness gets at an important idea in network analysis, that of
the flow of information or other traffic and of the influence vertices might
have over that flow. However, betweenness as defined by Freeman is based
on counting only the shortest paths between vertex pairs, effectively assuming
that all or at least most traffic passes along those shortest paths. In reality traf-

14
This score is shared by many actors. It is the minimum possible score of 2n − 1 as described
above.
15
Another possibility, proposed by Freeman [128] in his original paper on betweenness, is to
divide by the maximum possible value that betweenness can take on any network of size n, which,
as mentioned above, occurs for the central vertex in a star graph. The resulting expression for
between is then
1 nist
n −n+1 ∑
xi = 2 .
st gst
We, however, prefer Eq. (7.38), which we find easier to interpret, although the difference between
the two becomes small anyway in the limit of large n.

190
 7.7 | B ETWEENNESS CENTRALITY

fic flows along paths other than the shortest in many networks. Most of us, for
instance, will have had the experience of hearing news about one of our friends
not from that friend directly but from another mutual acquaintance—the mes-
sage has passed along a path of length two via the mutual acquaintance, rather
than along the direct (geodesic) path of length one.
A version of betweenness centrality that makes some allowance for effects
like this is the flow betweenness, which was proposed by Freeman et al. [130]
and is based on the idea of maximum flow. Imagine each edge in a network See Section 6.12 for a dis-
as a pipe that can carry a unit flow of some fluid. We can ask what the maxi- cussion of maximum flow
mum possible flow then is between a given source vertex s and target vertex t in networks.
through these pipes. In general the answer is that more than a single unit of
flow can be carried between source and target by making simultaneous use of
several different paths through the network. The flow betweenness of a ver-
tex i is defined according to Eq. (7.35), but with nist being now the amount of
flow through vertex i when the maximum flow is transmitted from s to t.
As we saw in Section 6.12, the maximum flow between vertices s and t
is also equal to the number of edge-independent paths between them. Thus
another way equivalent to look at the flow betweenness would be to consider
nist to be the number of independent paths between s and t that run through
vertex i.
A slight problem arises because the independent paths between
a given pair of vertices are not necessarily unique. For instance, A
the network shown in Fig. 7.3 has two edge-independent paths be-
tween s and t but we have two choices about what those paths are,
either the paths denoted by the solid arrows, or those denoted by s t
the dashed ones. Furthermore, our result for the flow betweenness
will depend on which choice we make; the vertices labeled A and B
B
fall on one set of paths but not the other. To get around this problem,
Freeman et al. define the flow through a vertex for their purposes to
be the maximum possible flow over all possible choices of paths, or
Figure 7.3: Edge-independent paths in
equivalently the maximum number of independent paths. Thus in a small network. The vertices s and t
the network of Fig. 7.3, the contribution of the flow between s and t in this network have two independent
to the betweenness of vertex A would be 1, since this is the maxi- paths between them, but there are two
mum value it takes over all possible choices of flow paths. distinct ways of choosing those paths,
In terms of our information analogy, one can think of flow be- represented by the solid and dashed
tweenness as measuring the betweenness of vertices in a network in curves.
which a maximal amount of information is continuously pumped
between all sources and targets. Flow betweenness takes account of more than
just the geodesic paths between vertices, since flow can go along non-geodesic
paths as well as geodesic ones. (For example, the paths through vertices A

191
M EASURES AND METRICS

and B in the example above are not geodesic.) Indeed, in some cases none of the
paths that appear in the solution of the maximum flow problem are geodesic
paths, so geodesic paths may not be counted at all by this measure.
But this point highlights a problem with flow betweenness: although it
typically counts more paths than the standard shortest-path betweenness, flow
betweenness still only counts a subset of possible paths, and some important
ones (such as geodesic paths) may be missed out altogether. One way to look at
the issue is that both shortest-path betweenness and flow betweenness assume
flows that are optimal in some sense—passing only along shortest paths in the
first case and maximizing total flow in the second. Just as there is no reason to
suppose that information or other traffic always takes the shortest path, there
is no reason in general to suppose it should act to maximize flow (although of
course there may be special cases in which it does).
A betweenness variant that does count all paths is the random-walk between-
ness [243]. In this variant traffic between vertices s and t is thought of as per-
See Section 6.14 for a dis- forming an (absorbing) random walk that starts at vertex s and continues un-
cussion of random walks. til it reaches vertex t. The betweenness is defined according to xi = ∑st nist
but with nist now being the number of times that the random walk from s to t
passes through i on its journey, averaged over many repetitions of the walk.
Note that in this case nist = nits in general, even on an undirected network.
For instance, consider this portion of a network:

s t

A random walk from s to t may pass through vertex A before returning to s and
stepping thence to t, but a walk from t to s will never pass through A because
its first step away from t will always take it to s and then the walk will finish.
Since every possible path from s to t occurs in a random walk with some
probability (albeit a very small one) the random-walk betweenness includes
contributions from all paths.16 Note, however, that different paths appear in
general with different probabilities, so paths do not contribute equally to the

16
All paths, that is, that terminate at the target vertex t the first time they reach it. Since we use
an absorbing random walk, paths that visit the target, move away again, and then return are not
included in the random-walk betweenness.

192
 7.8 | G ROUPS OF VERTICES

betweenness scores, longer paths typically making smaller contributions than

shorter ones, a bias that is plausible in some but by no means all cases.
Random walk betweenness would be an appropriate betweenness measure
for traffic that traverses a network with no idea of where it is going—it sim-
ply wanders around at random until it reaches its destination. Shortest-path
betweenness is the exact opposite. It is the appropriate measure for informa-
tion that knows exactly where it is going and takes the most direct path to get
there. It seems likely that most real-world situations fall somewhere in be-
tween these two extremes. However, it is found in practice [243] that the two
measures often give quite similar results, in which case one can with reason-
able justification assume that the “correct” answer, the one lying between the
limits set by the shortest-path and random-walk measures, is similar to both.
In cases where the two differ by a considerable margin, however, we should
be wary of attributing too much authority to either measure—there is no guar-
antee that either is telling us a great deal about true information flow in the
network.
Other generalizations of betweenness are also possible, based on other mod-
els of diffusion, transmission, or flow along network edges. We refer the inter-
ested reader to the article by Borgatti [51], which draws together many of the
possibilities into a broad general framework for betweenness measures.

7.8 G ROUPS OF VERTICES

Many networks, including social and other networks, divide naturally into
groups or communities. Networks of people divide into groups of friends,
coworkers, or business partners; the World Wide Web divides into groups of
related web pages; biochemical networks divide into functional modules, and
so forth. The definition and analysis of groups within networks is a large and
fruitful area of network theory. In Chapter 11 we discuss some of the sophisti-
cated computer methods that have been developed for detecting groups, such
as hierarchical clustering and spectral partitioning. In this section we discuss
some simpler concepts of network groups which can be useful for probing and
describing the local structure of networks. The primary constructs we look at
are cliques, plexes, cores, and components.

7.8.1 C LIQUES , PLEXES , AND CORES

A clique is a maximal subset of the vertices in an undirected network such that
every member of the set is connected by an edge to every other. The word
“maximal” here means that there is no other vertex in the network that can

193
M EASURES AND METRICS

be added to the subset while preserving the property that every vertex is con-
nected to every other. Thus a set of four vertices in a network would be a clique
if (and only if) each of the four is directly connected by edges to the other three
and if there is no other vertex anywhere in the network that could be added to
make a group of five vertices all connected to each other. Note that cliques can
overlap, meaning that they can share one or more of the same vertices.
The occurrence of a clique in an otherwise sparse network is normally an
indication of a highly cohesive subgroup. In a social network, for instance, one
might encounter a set of individuals each of whom was acquainted with each
of the others, and such a clique would probably indicate that the individuals
A clique of four vertices in question are closely connected—a set of coworkers in an office for example
within a network. or a group of classmates in a school.
However, it’s also the case that many circles of friends form only near-
cliques, rather than perfect cliques. There may be some members of the group
who are unacquainted, even if most members know one another. The require-
ment that every possible edge be present within a clique is a very stringent
A one, and it seems natural to consider how we might relax this requirement.
One construct that does this is the k-plex. A k-plex of size n is a maximal subset
of n vertices within a network such that each vertex is connected to at least
n − k of the others. If k = 1, we recover the definition of an ordinary clique—a
B
1-plex is the same as a clique. If k = 2, then each vertex must be connected
to all or all-but-one of the others. And so forth.17 Like cliques, k-plexes can
Two overlapping cliques. overlap one another; a single vertex can belong to more than one k-plex.
Vertices A and B in this net- The k-plex is a useful concept for discovering groups within networks: in
work both belong to two real life many groups in social and other networks form k-plexes. There is
cliques of four vertices. no solid rule about what value k should take. Experimentation starting from
small values is the usual way to proceed. Smaller values of k tend to be mean-
ingful for smaller groups, whereas in large groups the smaller values impose
too stringent a constraint but larger values often give useful results. This sug-
gests another possible generalization of the clique idea: one could specify that
each member be connected to a certain fraction of the others, say 75% or 50%.
(As far as we know, this variant doesn’t have a name and it is not in wide use,
but perhaps it should be.)
Many other variations on the clique idea have been proposed in the litera-
ture. For instance Flake et al. [122] proposed a definition of a group as a subset

17
This definition is slightly awkward to remember, since the members of a k-plex are allowed
to be unconnected to k − 1 other members and not k. It would perhaps have been more sensible to
define k such that a 0-plex was equivalent to a normal clique, but for better or worse we are stuck
with the definition we have.

194
 7.8 | G ROUPS OF VERTICES

of vertices such that each has at least as many connections to vertices inside the
group as to vertices outside. Radicchi et al. [276] proposed a weaker definition
of a group as a subset of vertices such that the total number of connections of
all vertices in the group to others in the group is greater than the total number
of connections to vertices outside.18
Another concept closely related to the k-plex is the k-core. A k-core is a
maximal subset of vertices such that each is connected to at least k others in
the subset.19 It should be obvious (or you can easily prove it for yourself) that
a k-core of n vertices is also an (n − k )-plex. However, the set of all k-cores
for a given value of k is not the same as the set of all k-plexes for any value
of k, since n, the size of the group, can vary from one k-core to another. Also,
unlike k-plexes (and cliques), k-cores cannot overlap, since by their definition
two k-cores that shared one or more vertices would just form a single larger
k-core.
The k-core is of particular interest in network analysis for the practical rea-
son that it is very easy to find the set of all k-cores in a network. A simple
algorithm is to start with your whole network and remove from it any vertices
that have degree less than k, since clearly such vertices cannot under any cir-
cumstances be members of a k-core. In so doing, one will normally also reduce
the degrees of some other vertices in the network—those that were connected
to the vertices just removed. So we then go through the network again to see
if there are any more vertices that now have degree less than k and if there are
we remove those too. And so we proceed, repeatedly pruning the network to
remove vertices with degree less than k until no such vertices remain.20 What
is left over will, by definition, be a k-core or a set of k-cores, since each vertex is
connected to at least k others. Note that we are not necessarily left with a single
k-core—there’s no guarantee that the network will be connected once we are
done pruning it, even if it was connected to start with.
Two other generalizations of cliques merit a brief mention. A k-clique is a
maximal subset of vertices such that each is no more than a distance k away
from any of the others via the edges of the network. For k = 1 this just recovers

18
Note that for the purposes of this latter definition, an edge between two vertices A and B
within the group counts as two connections, one from A to B and one from B to A.
19
We have to be careful about the meaning of the word “maximal” here. It is possible to have a
group of vertices such that each is connected to at least k others and no single vertex can be added
while retaining this property, but it may be possible to add more than one vertex. Such groups,
however, are not considered to be k-cores. A group is only a k-core if it is not a subset of any larger
group that is a k-core.
20
A closely related process, bootstrap percolation, has also been studied in statistical physics,
principally on regular lattices.

195
M EASURES AND METRICS

the definition of an ordinary clique. For larger k it constitutes a relaxation

of the stringent requirements of the usual clique definition. Unfortunately it
is not a very well-behaved one, since a k-clique by this definition need not
be connected via paths that run within the subset (see figure). If we restrict
ourselves to paths that run only within the subset then the resulting object is
known as either a k-clan or a k-club. (The difference between the two lies in
whether we impose the restriction that paths stay within the group from the
The outlined set of three outset, or whether we first find k-cliques and then discard those with outside
vertices in this network paths. The end results can be different in the two cases. For more details see
constitute a 2-clique, but Wasserman and Faust [320].).
one that is not connected
via paths within the 2-
clique.
7.8.2 C OMPONENTS AND k- COMPONENTS
In Section 6.11 we introduced the concept of a component. A component in an
undirected network is a maximal subset of vertices such that each is reachable
by some path from each of the others. A useful generalization of this concept
is the k-component. A k-component (sometimes also called a k-connected compo-
nent) is a maximal subset of vertices such that each is reachable from each of
the others by at least k vertex-independent paths—see Fig. 7.4. (Recall that two
paths are said to be vertex-independent if they share none of the same vertices,
except the starting and ending vertices—see Section 6.12.) For the common
special cases k = 2 and k = 3, k-components are also called bicomponents and
tricomponents respectively.
A 1-component by this definition is just an ordinary component—there is
at least one path between every pair of vertices—and k-components for k ≥ 2
are nested within each other. A 2-component or bicomponent, for example, is
necessarily a subset of a 1-component, since any pair of vertices that are con-
nected by at least two paths are also connected by at least one path. Similarly
a tricomponent is necessarily a subset of a bicomponent, and so forth. (See
Fig. 7.4 again.)
As discussed in Section 6.12, the number of vertex-independent paths be-
tween two vertices is equal to the size of the vertex cut set between the same
two vertices, i.e., the number of vertices that would have to be removed in or-
der to disconnect the two. So another way of defining a k-component would
be to say that it is a maximal subset of vertices such that no pair of vertices can
be disconnected from each other by removing less than k vertices.
A variant of the k-component can also be defined using edge-independent
paths, so that vertices are in the same k-component if they are connected by k or
more edge-independent paths, or equivalently if they cannot be disconnected
by the removal of less than k edges. In principal this variant could be useful in

196
 7.8 | G ROUPS OF VERTICES

1−component
2−component
3−component

Figure 7.4: The k-components in a small network. The shaded regions denote the k-
components in this small network, which has a single 1-component, two 2-components,
one 3-component, and no k-components for any higher value of k. Note that the
k-components are nested within one another, the 2-components falling inside the 1-
component and the 3-component falling inside one of the 2-components.

certain circumstances but in practice it is rarely used.

The idea of a k-component is a natural one in network analysis, being con-
nected with the idea of network robustness. For instance, in a data network
such as the Internet, the number of vertex-independent paths between two
vertices is also the number of independent routes that data might take be-
tween the same two vertices, and the size of the cut set between them is the
number of vertices in the network—typically routers—that would have to fail
or otherwise be knocked out to sever the data connection between the two end-
points. Thus a pair of vertices connected by two independent paths cannot be
disconnected from one another by the failure of any single router. A pair of
vertices connected by three paths cannot be disconnected by the failure of any
two routers. And so forth. A k-component with k ≥ 2 in a network like the
Internet is a subset of the network that has robust connectivity in this sense.
One would hope, for instance, that most of the network backbone—the system
of high volume world-spanning links that carry long-distance data (see Sec- The two highlighted ver-
tion 2.1)—is a k-component with high k, so that it would be difficult for points tices in this network form a
on the backbone to lose connection with one another. tricomponent, even though
Note that for k ≥ 3, the k-components in a network can be non-contiguous they are not directly con-
(see figure). Ordinary components (1-components) and bicomponents, by con- nected to each other. The
other three vertices are not
trast, are always contiguous. Within the social networks literature, where non-
in the tricomponent.
contiguous components are often considered undesirable, k-components are

197
M EASURES AND METRICS

sometimes defined slightly differently: a k-component is defined to be a max-

imal subset of vertices such that every pair in the set is connected by at least k
vertex-independent paths that themselves are contained entirely within the subset.
This definition rules out non-contiguous k-components, but it is also mathe-
matically and computationally more difficult to work with than the standard
definition. For this reason, and because there are also plenty of cases in which
it is appropriate to count non-contiguous k-components, the standard defini-
tion remains the most widely used one in fields other than sociology.

7.9 T RANSITIVITY
A property very important in social networks, and useful to a lesser degree in
other networks too, is transitivity. In mathematics a relation “◦” is said to be
transitive if a ◦ b and b ◦ c together imply a ◦ c. An example would be equality.
If a = b and b = c, then it follows that a = c also, so “=” is a transitive relation.
Other examples are “greater than,” “less than,” and “implies.”
In a network there are various relations between pairs of vertices, the sim-
plest of which is “connected by an edge.” If the “connected by an edge” re-
lation were transitive it would mean that if vertex u is connected to vertex v,
and v is connected to w, then u is also connected to w. In common parlance,
“the friend of my friend is also my friend.” Although this is only one possi-
ble kind of network transitivity—other network relations could be transitive
too—it is the only one that is commonly considered, and networks showing
this property are themselves said to be transitive. This definition of network
transitivity could apply to either directed or undirected networks, but let us
take the undirected case first, since it’s simpler.
Perfect transitivity only occurs in networks where each component is a
fully connected subgraph or clique, i.e., a subgraph in which all vertices are
connected to all others.21 Perfect transitivity is therefore pretty much a useless
concept in networks. However, partial transitivity can be very useful. In many
networks, particularly social networks, the fact that u knows v and v knows w

21
To see this suppose we have a component that is perfectly transitive but not a clique, i.e., there
is at least one pair of vertices u, w in the component that are not directly connected by an edge.
Since u and w are in the same component they must therefore be connected by some path of length
greater than one, u, v1 , v2 , v3 , . . . , w. Consider the first two links in this path. Since u is connected
by an edge to v1 and v1 to v2 it follows that u must be connected to v2 if the network is perfectly
transitive. Then consider the next two links. Since u is connected to v2 and v2 to v3 it follows that
u must be connected to v3 . Repeating the argument all the way along the path, we can then see
that u must be connected by an edge to w. But this violates the hypothesis that u and w are not
directly connected. Hence no perfectly transitive components exist that are not cliques.

198
 7.9 | T RANSITIVITY

doesn’t guarantee that u knows w, but makes it much more likely. The friend
of my friend is not necessarily my friend, but is far more likely to be my friend
than some randomly chosen member of the population.
We can quantify the level of transitivity in a network as follows. If u knows u w
v and v knows w, then we have a path uvw of two edges in the network. If u
also knows w, we say that the path is closed—it forms a loop of length three,
or a triangle, in the network. In the social network jargon, u, v, and w are said v
to form a closed triad. We define the clustering coefficient22 to be the fraction of The path uvw (solid edges)
paths of length two in the network that are closed. That is, we count all paths is said to be closed if the
of length two, and we count how many of them are closed, and we divide the third edge directly from u
second number by the first to get a clustering coefficient C that lies in the range to w is present (dashed
from zero to one: edge).

(number of closed paths of length two)

C= . (7.39)
(number of paths of length two)

C = 1 implies perfect transitivity, i.e., a network whose components are all

cliques. C = 0 implies no closed triads, which happens for various topologies,
such as a tree (which has no closed loops of any kind—see Section 6.7) or a
square lattice (which has closed loops with even numbers of vertices only and
no closed triads).
Note that paths in networks, as defined in Section 6.10 have a direction and
two paths that traverse the same edges but in opposite directions are counted
separately in Eq. (7.39). Thus uvw and wvu are distinct paths and are counted
separately. Similarly, closed paths are counted separately in each direction.23
An alternative way to write the clustering coefficient is

(number of triangles) × 6
C= . (7.40)
(number of paths of length two)

Why the factor of six? It arises because each triangle in the network gets
counted six times over when we count up the number of closed paths of length
two. Suppose we have a triangle uvw. Then there are six paths of length two

22
It’s not entirely clear why the clustering coefficient has the name it has. The name doesn’t
appear to be connected with the earlier use of the word clustering in social network analysis to
describe groups or clusters of vertices (see Section 11.11.2). The reader should be careful to avoid
confusing these two uses of the word.
23
In fact, we could count each path just in one direction, provided we did it for both the nu-
merator and denominator of Eq. (7.39). Doing so would decrease both counts by a factor of two,
but the factors would cancel and the end result would be the same. In most cases, and particularly
when writing computer programs, it is easier to count paths in both directions—it avoids having
to remember which paths you have counted before.

199
M EASURES AND METRICS

in it: uvw, vwu, wuv, wvu, vuw, and uwv. Each of these six is closed, so the
number of closed paths is six times the number of triangles.
Yet another way to write the clustering coefficient would be to note that
if we have a path of length two, uvw, then it is also true to say that vertices
u and w have a common neighbor in v—they share a mutual acquaintance
in social network terms. If the triad uvw is closed then u and w are them-
selves acquainted, so the clustering coefficient can be thought of also as the
fraction of pairs of people with a common friend who are themselves friends
or equivalently as the mean probability that two people with a common friend
are themselves friends. This is perhaps the most common way of defining the
clustering coefficient. In mathematical notation:
(number of triangles) × 3
C= . (7.41)
(number of connected triples)

Here a “connected triple” means three vertices uvw with edges (u, v) and (v, w).
(The edge (u, w) can be present or not.) The factor of three in the numerator
arises because each triangle gets counted three times when we count the con-
nected triples in the network. The triangle uvw for instance contains the triples
A triangle contains six dis-
uvw, vwu, and wuv. In the older social networks literature the clustering coef-
tinct paths of length two, ficient is sometimes referred to as the “fraction of transitive triples,” which is a
all of them closed. reference to this definition of the coefficient.
Social networks tend to have quite high values of the clustering coefficient.
For example, the network of film actor collaborations discussed earlier has
been found to have C = 0.20 [241]; a network of collaborations between bi-
ologists has been found to have C = 0.09 [236]; a network of who sends email
to whom in a large university has C = 0.16 [103]. These are typical values for
social networks. Some denser networks have even higher values, as high as 0.5
or 0.6. (Technological and biological networks by contrast tend to have some-
what lower values. The Internet at the autonomous system level, for instance,
has a clustering coefficient of only about 0.01. This point is discussed in more
detail in Section 8.6.)
In what sense are these clustering coefficients for social networks high?
Well, let us assume, to make things simple, that everyone in a network has
about the same number c of friends. Consider one of my friends in this net-
work and suppose they pick their friends completely at random from the whole
population. Then the chance that one of their c friends happens to be a partic-
ular one of my other friends would be c/n, where n is the size of the network.
Thus in this network the probability of two of my friends being acquainted,
which is by definition the clustering coefficient, would be just c/n. Of course
it is not the case that everyone in a network has the same number of friends,

200
 7.9 | T RANSITIVITY

and we will see how to perform better calculations of the clustering coefficient
later (Section 13.4), but this crude calculation will serve our purposes for the
moment.
For the networks cited above, the value of c/n is 0.0003 (film actors), 0.00001
(biology collaborations), and 0.00002 (email messages). Thus the measured
clustering coefficients are much larger than this estimate based on the assump-
tion of random network connections. Even though the estimate ignores, as
we have said, any variation in the number of friends people have, the dispar-
ity between the calculated and observed values of the clustering coefficient
is so large that it seems unlikely it could be eliminated just by allowing the
number of friends to vary. A much more likely explanation is that our other
assumption, that people pick their friends at random, is seriously flawed. The
numbers suggest that there is a much greater chance that two people will be
acquainted if they have another common acquaintance than if they don’t.
Although this argument is admittedly crude, we will see in Section 8.6 how
to make it more accurate and so show that our basic conclusion is indeed cor-
rect.
Some social networks, such as the email network above, are directed net-
works. In calculating clustering coefficients for direct networks, scientists have
typically just ignored their directed nature and applied Eq. (7.41) as if the edges u w
were undirected. It is however possible to generalize transitivity to take ac-
count of directed links. If we have a directed relation between vertices such
as “u likes v” then we can say that a triple of vertices is closed or transitive if
u likes v, v likes w, and also u likes w. (Note that there are many distinct ways v
for such a triple to be transitive, depending on the directions of the edges. The A transitive triple of ver-
example given here is only one of six different possibilities.) One can calculate tices in a directed network.
a clustering coefficient or fraction of transitive triples in the obvious fashion for
the directed case, counting all directed paths of length two that are closed and
dividing by the total number of directed paths of length two. For some reason,
however, such measurements have not often appeared in the literature.

7.9.1 L OCAL CLUSTERING AND REDUNDANCY

We can also define a clustering coefficient for a single vertex. For a vertex i, we
define24
(number of pairs of neighbors of i that are connected)
Ci = . (7.42)
(number of pairs of neighbors of i)

24
The notation Ci is used for both the local clustering coefficient and the closeness centrality
and we should be careful not to confuse the two.

201
M EASURES AND METRICS

That is, to calculate Ci we go through all distinct pairs of vertices that are neigh-
bors of i in the network, count the number of such pairs that are connected to
each other, and divide by the total number of pairs, which is 12 k i (k i − 1) where
k i is the degree of i,. Ci is sometimes called the local clustering coefficient and it
represents the average probability that a pair of i’s friends are friends of one
another.
Local clustering is interesting for several reasons. First, in many networks
Structural holes it is found empirically to have a rough dependence on degree, vertices with
higher degree having a lower local clustering coefficient on average. This point
is discussed in detail in Section 8.6.1.
Second, local clustering can be used as a probe for the existence of so-called
“structural holes” in a network. While it is common in many networks, es-
pecially social networks, for the neighbors of a vertex to be connected among
themselves, it happens sometimes that these expected connections between
neighbors are missing. The missing links are called structural holes and were
first studied in this context by Burt [60]. If we are interested in efficient spread
When the neighbors of a of information or other traffic around a network, as we were in Section 7.7,
node are not connected to then structural holes are a bad thing—they reduce the number of alternative
one another we say the net- routes information can take through the network. On the other hand structural
work contains “structural holes can be a good thing for the central vertex i whose friends lack connec-
holes.” tions, because they give i power over information flow between those friends.
If two friends of i are not connected directly and their information about one
another comes instead via their mutual connection with i then i can control the
flow of that information. The local clustering coefficient measures how influ-
ential i is in this sense, taking lower values the more structural holes there are
in the network around i. Thus local clustering can be regarded as a type of
centrality measure, albeit one that takes small values for powerful individuals
rather than large ones.
In this sense, local clustering can also be thought of as akin to the between-
ness centrality of Section 7.7. Where betweenness measures a vertex’s control
over information flowing between all pairs of vertices in its component, lo-
cal clustering is like a local version of betweenness that measures control over
flows between just the immediate neighbors of a vertex. One measure is not
necessarily better than another. There may be cases in which we want to take
all vertices into account and others where we want to consider only immedi-
ate neighbors—the choice will depend on the particular questions we want to
answer. It is worth pointing out however that betweenness is much more com-
putationally intensive to calculate than local clustering (see Section 10.3.6), and
that in practice betweenness and local clustering are strongly correlated [60].
There may in many cases be little to be gained by performing the more costly

202
 7.9 | T RANSITIVITY

full calculation of betweenness and much to be saved by sticking with cluster-

ing, given that the two contain much the same information.25
In his original studies of structural holes, Burt [60] did not in fact
make use of the local clustering coefficient as a measure of the pres-
ence of holes.26 Instead, he used another measure, which he called
redundancy. The original definition of redundancy was rather com-
plicated, but Borgatti [50] has shown that it can be simplified to the
following: the redundancy Ri of a vertex i is the mean number of
connections from a neighbor of i to other neighbors of i. Consider
the example shown in Fig. 7.5 in which vertex i has four neighbors. Figure 7.5: Redundancy. The neigh-
Each of those four could be acquainted with any of the three others, bors of the central vertex in this fig-
but in this case none of them is connected to all three. One is con- ure have 0, 1, 1, and 2 connections to
nected to none of the others, two are connected to one other, and other neighbors respectively. The re-
the last is connected to two others. The redundancy is the average dundancy is the mean of these values:
of these numbers Ri = 14 (0 + 1 + 1 + 2) = 1. The minimum possi- Ri = 14 (0 + 1 + 1 + 2) = 1.
ble value of the redundancy of a vertex is zero and the maximum is
k i − 1, where k i is the degree of vertex i.
It’s probably obvious that Ri is related to the local clustering Ci . To see
precisely what the relation is, we note that if the average number of connec-
tions from a friend of i to other friends is Ri , then the total number of connec-
tions between friends is 12 k i Ri . And the total number of pairs of friends of i is
2 k i ( k i − 1). The local clustering coefficient, Eq. (7.42), is the ratio of these two
1

quantities:
1
k i Ri Ri
Ci = 1 2 = . (7.43)
2 k i ( k i − 1)
k i −1

Given that k i − 1 is the maximum value of Ri , the local clustering coefficient

can be thought of as simply a version of the redundancy rescaled to have a
maximum value of 1. Applying Eq. (7.43) to the example of Fig. 7.5 implies
that the local clustering coefficient for the central vertex should be Ci = 13 , and
the reader can easily verify that this is indeed the case.
A third context in which the local clustering coefficient arises is in the cal-
culation of the global clustering coefficient itself. Watts and Strogatz [323] pro-
posed calculating a clustering coefficient for an entire network as the mean of

25
As an example, in Section 11.11.1 we study methods for partitioning networks into clusters
or communities and we will see that effective computer algorithms for this task can be created
based on betweenness measures, but that almost equally effective and much faster algorithms can
be created based on local clustering.
26
Actually, the local clustering coefficient hadn’t yet been invented. It was first proposed to this
author’s knowledge by Watts [321] a few years later.

203
M EASURES AND METRICS

the local clustering coefficients for each vertex:

n
1
CWS =
n ∑ Ci , (7.44)
i =1

where n is the number of vertices in the network. This is a different defini-

tion for the clustering coefficient from the one given earlier, Eq. (7.41), and
the two definitions are not equivalent. Furthermore, they can give substan-
tially different numbers for a given network and because both definitions are
in reasonably common use this can give rise to confusion. We favor our first
definition for C, Eq. (7.41), because it has a simple interpretation and because
it is normally easier to calculate. Also the second definition, Eq. (7.44), tends
to be dominated by vertices with low degree, since they have small denom-
inators in Eq. (7.42), and the measure thus gives a rather poor picture of the
overall properties of any network with a significant number of such vertices.27
It’s worth noting, however, that the definition of Eq. (7.44) was actually pro-
posed before Eq. (7.41) and, perhaps because of this, it finds moderately wide
use in network studies. So you need at least to be aware of both definitions
and clear which is being used in any particular situation.

7.10 R ECIPROCITY
The clustering coefficient of Section 7.9 measures the frequency with which
loops of length three—triangles—appear in a network. Of course, there is
no reason why one should concentrate only on loops of length three, and
people have occasionally looked at the frequency of loops of length four or
more [44,61,133,140,238]. Triangles occupy a special place however because in
an undirected simple graph the triangle is the shortest loop we can have (and
usually the most commonly occurring). However, in a directed network this is
not the case. In a directed network, we can have loops of length two—a pair of
vertices between which there are directed edges running in both directions—
and it is interesting to ask about the frequency of occurrence of these loops
also.
A loop of length two in a The frequency of loops of length two is measured by the reciprocity, and tells
directed network. you how likely it is that a vertex that you point to also points back at you. For
instance, on the World Wide Web if my web page links to your web page, how
likely is it, on average, that yours link back again to mine? In general, it’s found

27
As discussed in Section 8.6.1, vertices with low degree tend to have high values of Ci in most
networks and this means that CWS is usually larger than the value given by Eq. (7.41), sometimes
much larger.

204
 7.10 | R ECIPROCITY

that you are much more likely to link to me if I link to you than if I don’t. (That
probably isn’t an Earth-shattering surprise, but it’s good to know when the
data bear out one’s intuitions.) Similarly in friendship networks, such as the
networks of schoolchildren described in Section 3.2 where respondents were
asked to name their friends, it is much more likely that you will name me if I
name you than if I do not.
If there is a directed edge from vertex i to vertex j in a directed network and
there is also an edge from j to i then we say the edge from i to j is reciprocated.
(Obviously the edge from j to i is also reciprocated.) Pairs of edges like this are
also sometimes called co-links, particularly in the context of the World Wide
Web [104].
The reciprocity r is defined as the fraction of edges that are reciprocated.
Noting that the product of adjacency matrix elements Aij A ji is 1 if and only if
there is an edge from i to j and an edge from j to i and is zero otherwise, we
can sum over all vertex pairs i, j to get an expression for the reciprocity:

1 1
r=
m ∑ Aij A ji = m
Tr A2 , (7.45)
ij

where m is, as usual, the total number of (directed) edges in the network.
Consider for example this small network of four vertices:

There are seven directed edges in this network and four of them are recipro-
cated, so the reciprocity is r = 47
0.57. In fact, this is about the same value as
seen on the World Wide Web. There is about a 57% percent chance that if web
page A links to web page B then B also links back to A.28 As another example,
in a study of a network of who has whom in their email address book it was
found that the reciprocity was about r = 0.23 [248].

28
This figure is an unusually high one among directed networks, but there are reasons for it.
One is that many of the links between web pages are between pages on the same website, and it is
common for such pages to link to each other. If you exclude links between pages on the same site
the value of the reciprocity is lower.

205
M EASURES AND METRICS

7.11 S IGNED EDGES AND STRUCTURAL BALANCE

In some social networks, and occasionally in other networks, edges are al-
lowed to be either “positive” or “negative.” For instance, in an acquaintance
network we could denote friendship by a positive edge and animosity by a
negative edge:

Enemies
Friends
One could also consider varying degrees of friendship or animosity—networks
with more strongly positive or negative edges in them—but for the moment
let’s stick to the simple case where each edge is in just one of two states, pos-
itive or negative, like or dislike. Such networks are called signed networks and
their edges are called signed edges.
It is important to be clear here that a negative edge is not the same as the
absence of an edge. A negative edge indicates, for example, two people who
interact regularly but dislike each other. The absence of an edge represents two
people who do not interact. Whether they would like one another if they did
interact is not recorded.
Now consider the possible configurations of three edges in a triangle in a
signed network, as depicted in Fig. 7.6. If “+” and “−” represent like and
dislike, then we can imagine some of these configurations creating social prob-
lems if they were to arise between three people in the real world. Configura-
tion (a) is fine: everyone likes everyone else. Configuration (b) is probably also
fine, although the situation is more subtle than (a). Individuals u and v like one
another and both dislike w, but the configuration can still be regarded as sta-
ble in the sense that u and v can agree over their dislike of w and get along just
fine, while w hates both of them. No one is conflicted about their allegiances.
Put another way, w is u’s enemy and v is w’s enemy, but there is no problem
with u and v being friends if one considers that the “enemy of my enemy is my
friend.”
Configuration (c) however could be problematic. Individual u likes indi-
vidual v and v likes w, but u thinks w is an idiot. This is going to place a strain
on the friendship between u and v because u thinks v’s friend is an idiot. Alter-
natively, from the point of view of v, v has two friends, u and w and they don’t
get along, which puts v in an awkward position. In many real-life situations
of this kind the tension would be resolved by one of the acquaintances being

206
 7.11 | S IGNED EDGES AND STRUCTURAL BALANCE

w w w w

u v u v u v u v

(a) (b) (c) (d)

Figure 7.6: Possible triad configurations in a signed network. Configurations (a) and
(b) are balanced and hence relatively stable, but configurations (c) and (d) are unbal-
anced and liable to break apart.

broken, i.e., the edge would be removed altogether. Perhaps v would simply
stop talking to one of his friends, for instance.
Configuration (d) is somewhat ambiguous. On the one hand, it consists
of three people who all dislike each other, so no one is in doubt about where
things stand: everyone just hates everyone else. On the other hand, the “en-
emy of my enemy” rule does not apply here. Individuals u and v might like to
form an alliance in recognition of their joint dislike of w, but find it difficult to
do so because they also dislike each other. In some circumstances this might
cause tension. (Think of the uneasy alliance of the US and Russia against Ger-
many during World War II, for instance.) But what one can say definitely is
that configuration (d) is often unstable. There may be little reason for the three
to stay together when none of them likes the others. Quite probably three ene-
mies such as these would simply sever their connections and go their separate
ways.
The feature that distinguishes the two stable configurations in Fig. 7.6 from
the unstable ones is that they have an even number of minus signs around the
loop.29 One can enumerate similar configurations for longer loops, of length
four or greater, and again find that loops with even numbers of minus signs
appear stable and those with odd numbers unstable.
This alone would be an observation of only slight interest, where it not Two stable configurations
for the intriguing fact that this type of stability really does appear have an in loops of length four.
effect on the structure of networks. In surveys it is found that the unstable
configurations in Fig. 7.6, the ones with odd numbers of minus signs, occur

29
This is similar in spirit to the concept of “frustration” that arises in the physics of magnetic
spin systems.

207
M EASURES AND METRICS

far less often in real social networks than the stable configurations with even
numbers of minus signs.
Networks containing only loops with even numbers of minus signs are
said to show structural balance, or sometimes just balance. An important conse-
quence of balance in networks was proved by Harary [154]:
A balanced network can be divided into connected groups of vertices
such that all connections between members of the same group are
positive and all connections between members of different groups are
negative.
Note that the groups in question can consist of a single vertex or many vertices,
and there may be only one group or there may be very many. Figure 7.7 shows
a balanced network and its division into groups. Networks that can be divided
into groups like this are said to be clusterable. Harary’s theorem tells us that all
balanced networks are clusterable.
Harary’s theorem is straightforward to prove, and the proof is
“constructive,” meaning that it shows not only when a network is
clusterable but also tells us what the groups are.30 We consider ini-
tially only networks that are connected—they have just one compo-
nent. In a moment we will relax this condition. We will color in the
vertices of the network each in one of two colors, denoted by the
open and filled circles in Fig. 7.7, for instance. We start with any
vertex we please and color it with whichever color we please. Then
Figure 7.7: A balanced, clusterable
we color in the others according to the following algorithm:
network. Every loop in this network
contains an even number of minus 1. A vertex v connected by a positive edge to another u that has
signs. The dotted lines indicate the di- already been colored gets colored the same as u.
vision of the network into clusters such 2. A vertex v connected by a negative edge to another u that has
that all acquaintances within clusters already been colored gets colored the opposite color from u.
have positive connections and all ac- For most networks it will happen in the course of this coloring pro-
quaintances in different clusters have cess that we sometimes come upon a vertex whose color has already
negative connections. been assigned. When this happens there is the possibility of a con-
flict arising between the previously assigned color and the one that
we would like to assign to it now according to the rules above. However, as
we now show, this conflict only arises if the network as a whole is unbalanced.
If in coloring in a network we come upon a vertex that has already been
colored in, it immediately implies that there must be another path by which
that vertex can be reached from our starting point and hence that there is at
least one, and possibly more than one, loop in the network to which this ver-

30
The proof we give is not Harary’s proof, which was quite different and not constructive.

208
 7.11 | S IGNED EDGES AND STRUCTURAL BALANCE

Even number of
Odd number of
minus signs

minus signs
u u

v v

(a) (b)

Figure 7.8: Proof that a balanced network is clusterable. If we fail to color a network in
two colors as described in the text, then there must exist a loop in the network that has
one or other of the two configurations shown here, both of which have an odd number
of minus signs around them (counting the one between the vertices u and v), and hence
the network is not balanced.

tex belongs—the loop consisting of the two paths between the starting point
and the vertex. Since the network is balanced, every loop to which our ver-
tex belongs must have an even number of negative edges around it. Now let
us suppose that the color already assigned to the vertex is in conflict with the
one we would like to assign it now. There are two ways in which this could
happen, as illustrated in Fig. 7.8. In case (a), we color in a vertex u and then
move onto its neighbor v, only to find that v has already been colored the op-
posite color to u, even though the edge between them is positive. This presents
a problem. But if u and v are opposite colors, then around any loop contain-
ing them both there must be an odd number of minus signs, so that the color
changes an odd number of times and ends up the opposite of what it started
out as. And if there is an odd number of minus signs around the loop, then the
network is not balanced.
In case (b) vertices u and v have the same color but the edge between them
is negative. Again we have a problem. But if u and v are the same color then
there must be an even number of negative edges around the rest of the loop
connecting them which, along with the negative edge between u and v, gives
us again an odd total number of negative edges around the entire loop, and
hence the network is again not balanced.
Either way, if we ever encounter a conflict about what color a vertex should
have then the network must be unbalanced. If the network is balanced, there-
fore, we will never encounter such a conflict and we will be able to color the
entire network with just two colors while obeying the rules.
Once we have colored the network in this way, we can immediately deduce
the identity of the groups that satisfy Harary’s theorem: we simply divide

209
M EASURES AND METRICS

the network into contiguous clusters of vertices that have the same color—see
Fig. 7.7 again. In every such cluster, since all vertices have the same color,
they must be joined by positive edges. Conversely, all edges that connected
different clusters must be negative, since the clusters have different colors. (If
they did not have different colors they would be considered the same cluster.)
Thus Harary’s theorem is proved and at the same time we have deduced a
method for constructing the clusters.31 It only remains to extend the proof to
networks that have more than one component, but this is trivial, since we can
simply repeat the proof above for each component separately.
The practical importance of Harary’s result rests on the fact that, as men-
tioned earlier, many real social networks are found naturally to be in a bal-
anced or mostly balanced state. In such cases it would be possible, therefore,
for the network to form into groups such that everyone likes others within
their group with whom they have contact and dislikes those in other groups.
It is widely assumed in social network theory that this does indeed often hap-
pen. Structural balance and clusterability in networks are thus a model for
cliquishness or insularity, with people tending to stick together in like-minded
groups and disdaining everyone outside their immediate community.
It is worth asking whether the inverse of Harary’s clusterability theorem
is also true. Is it also the case that a network that is clusterable is necessarily
balanced? The answer is no, as this simple counter-example shows:

31
As an interesting historical note, we observe that while Harary’s proof of his theorem is per-
fectly correct, his interpretation of it was, in this author’s opinion, erroneous. In his 1953 pa-
per [154], he describes the meaning of the theorem in the following words: “A psychological in-
terpretation of Theorem 1 is that a ‘balanced group’ consists of two highly cohesive cliques which
dislike each other.” (Harary is using the word “clique” in a non-technical sense here to mean a
closed group of people, rather than in the graph theoretical sense of Section 7.8.1.) However, just
because it is possible to color the network in two colors as described above does not mean the net-
work forms two groups. Since the vertices of a single color are not necessarily contiguous, there
are in general many groups of each color, and it seems unreasonable to describe these groups as
forming a single “highly cohesive clique” when in fact they have no contact at all. Moreover, it is
neither possible nor correct to conclude that the members of two groups of opposite colors dislike
each other unless there is at least one edge connecting the two. If two groups of opposite colors
never actually have any contact then it might be that they would get along just fine if they met.
It’s straightforward to prove that such an occurrence would lead to an unbalanced network, but
Harary’s statement says that the present balanced network implies dislike, and this is untrue. Only
if the network were to remain balanced upon addition of one or more edges between groups of
unlike colors would his conclusion be accurate.

210
 7.12 | S IMILARITY

In this network all three vertices dislike each other, so there is an odd number
of minus signs around the loop, but there is no problem dividing the network
into three clusters of one vertex each such that everyone dislikes the members
of the other clusters. This network is clusterable but not balanced.

7.12 S IMILARITY
Another central concept in social network analysis is that of similarity between
vertices. In what ways can vertices in a network be similar, and how can we
quantify that similarity? Which vertices in a given network are most similar
to one another? Which vertex v is most similar to a given vertex u? Answers
to questions like these can help us tease apart the types and relationships of
vertices in social networks, information networks, and others. For instance,
one could imagine that it might be useful to have a list of web pages that are
similar—in some appropriate sense—to another page that we specify. In fact,
several web search engines already provide a feature like this: “Click here for
pages similar to this one.”
Similarity can be determined in many different ways and most of them
have nothing to do with networks. For example, commercial dating and match-
making services try to match people with others to whom they are similar by
using descriptions of people’s interests, background, likes, and dislikes. In ef-
fect, these services are computing similarity measures between people based
on personal characteristics. Our focus in this book, however, is on networks,
so we will concentrate on the more limited problem of determining similar-
ity between the vertices of a network using the information contained in the
network structure.
There are two fundamental approaches to constructing measures of net-
work similarity, called structural equivalence and regular equivalence. The names
are rather opaque, but the ideas they represent are simple enough. Two ver-
tices in a network are structurally equivalent if they share many of the same
network neighbors. In Fig. 7.9a we show a sketch depicting structural equiv-
alence between two vertices i and j—the two share, in this case, three of the
same neighbors, although both also have other neighbors that are not shared.
Regular equivalence is more subtle. Two regularly equivalent vertices do
not necessarily share the same neighbors, but they have neighbors who are

211
M EASURES AND METRICS

i j i j

(a) Structural equivalence (b) Regular equivalence

Figure 7.9: Structural equivalence and regular equivalence. (a) Vertices i and j are
structurally equivalent if they share many of the same neighbors. (b) Vertices i and j
are regularly equivalent if their neighbors are themselves equivalent (indicated here by
the different shades of vertices).

themselves similar. Two history students at different universities, for example,

may not have any friends in common, but they can still be similar in the sense
that they both know a lot of other history students, history instructors, and so
forth. Similarly, two CEOs at two different companies may have no colleagues
in common, but they are similar in the sense that they have professional ties to
their respective CFO, CIO, members of the board, company president, and so
forth. Regular equivalence is illustrated in Fig. 7.9b.
In the next few sections we describe some mathematical measures that
quantify these ideas of similarity. As we will see, measures for structural
equivalence are considerably better developed than those for regular equiv-
alence.

7.12.1 C OSINE SIMILARITY

We start by looking at measures of structural equivalence and we will concen-
trate on undirected networks. Perhaps the simplest and most obvious mea-
sure of structural equivalence would be just a count of the number of common
neighbors two vertices have. In an undirected network the number nij of com-
mon neighbors of vertices i and j is given by

nij = ∑ Aik Akj , (7.46)

k

which is the ijth element of A2 . This quantity is closely related to the “co-
citation” measure introduced in Section 6.4.1. Cocitation is defined for directed

212
 7.12 | S IMILARITY

networks whereas we are here considering undirected ones, but otherwise it is

essentially the same thing.
However, a simple count of common neighbors for two vertices is not on
its own a very good measure of similarity. If two vertices have three common
neighbors is that a lot or a little? It’s hard to tell unless we know, for instance,
what the degrees of the vertices are, or how many common neighbors other
pairs of vertices share. What we need is some sort of normalization that places
the similarity value on some easily understood scale. One strategy might be
simply to divide by the total number of vertices in the network n, since this is
the maximum number of common neighbors two vertices can have in a simple
graph. (Technically the maximum is actually n − 2, but the difference is small
when n is large.) However, this unduly penalizes vertices with low degree: if
a vertex has degree three, then it can have at most three neighbors in common
with another vertex, but the two vertices would still receive a small similarity
value if the divisor n were very large. A better measure would allow for the
varying degrees of vertices. Such a measure is the cosine similarity, sometimes
also called Salton’s cosine.
In geometry, the inner or dot product of two vectors x and y is given by
x · y = |x| |y| cos θ, where |x| is the magnitude of x and θ is the angle between
the two vectors. Rearranging, we can write the cosine of the angle as
x·y
cos θ = . (7.47)
|x| |y|

Salton [290] proposed that we regard the ith and jth rows (or columns) of the
adjacency matrix as two vectors and use the cosine of the angle between them
as our similarity measure. Noting that the dot product of two rows is simply
∑k Aik Akj for an undirected network, this gives us a similarity

∑k Aik Akj
σij = cos θ =   . (7.48)
∑k A2ik ∑k A2jk

Assuming our network is an unweighted simple graph, the elements of the

adjacency matrix take only the values 0 and 1, so that A2ij = Aij for all i, j. Then
∑k A2ik = ∑k Aik = k i , where k i is the degree of vertex i (see Eq. (6.19)). Thus

∑k Aik Akj nij

σij = √ = √ . (7.49)
ki k j ki k j
The cosine similarity of i and j is therefore the number of common neighbors
of the two vertices divided by the geometric mean of their degrees. For the

213
M EASURES AND METRICS

vertices i and j depicted in Fig. 7.9a, for instance, the cosine similarity would
be
3
σij = √ = 0.671 . . . (7.50)
4×5
Notice that the cosine similarity is technically undefined if one or both of the
vertices has degree zero, but by convention we normally say in that case that
σij = 0.
The cosine similarity provides a natural scale for our similarity measure.
Its value always lies in the range from 0 to 1. A cosine similarity of 1 indicates
that two vertices have exactly the same neighbors. A cosine similarity of zero
indicates that they have none of the same neighbors. Notice that the cosine
similarity can never be negative, being a sum of positive terms, even though
cosines in general can of course be negative.

7.12.2 P EARSON COEFFICIENTS

An alternative way to normalize the count of common neighbors is to compare
it with the expected value that count would take on a network in which ver-
tices choose their neighbors at random. This line of argument leads us to the
Pearson correlation coefficient.
Suppose vertices i and j have degrees k i and k j respectively. How many
common neighbors should we expect them to have? This is straightforward to
calculate if they choose their neighbors purely at random. Imagine that vertex i
chooses k i neighbors uniformly at random from the n possibilities open to it (or
n − 1 on a network without self-loops, but the distinction is slight for a large
network), and vertex j similarly chooses k j neighbors at random. For the first
neighbor that j chooses there is a probability of k i /n that it will choose one of
the ones k i chose, and similarly for each succeeding choice. (We neglect the
possibility of choosing the same neighbor twice, since it is small for a large
network.) Then in total the expected number of common neighbors between
the two vertices will be k j times this, or k i k j /n.
A reasonable measure of similarity between two vertices is the actual num-
ber of common neighbors they have minus the expected number that they

214
 7.12 | S IMILARITY

would have if they chose their neighbors at random:

ki k j 1
∑ Aik A jk − n
= ∑ Aik A jk − n ∑ Aik ∑ A jl
k k k l

= ∑ Aik A jk − n Ai  A j
k

= ∑ Aik A jk −  Ai  A j
k

= ∑( Aik −  Ai )( A jk −  A j ), (7.51)
k

where  Ai denotes the mean n−1 ∑k Aik of the elements of the ith row of the
adjacency matrix. Equation (7.51) will be zero if the number of common neigh-
bors of i and j is exactly what we would expect on the basis of random chance.
If it is positive, then i and j have more neighbors than we would expect by
chance, which we take as an indication of similarity between the two. Equa-
tion (7.51) can also be negative, indicating that i and j have fewer neighbors
than we would expect, a possible sign of dissimilarity.
Equation (7.51) is simply n times the covariance cov( Ai , A j ) of the two rows
of the adjacency matrix. It is common to normalize the covariance, as we did
with the cosine similarity, so that its maximum value is 1. The maximum value
of the covariance of any two sets of quantities occurs when the sets are exactly
the same, in which case their covariance is just equal to the variance of either
set, which we could write as σi2 or σj2 , or in symmetric form as σi σj . Normaliz-
ing by this quantity then gives us the standard Pearson correlation coefficient:
cov( Ai , A j ) ∑k ( Aik −  Ai )( A jk −  A j )
rij = =   . (7.52)
σi σj ∑k ( Aik −  Ai )2 ∑k ( A jk −  A j )2

This quantity lies strictly in the range −1 ≤ rij ≤ 1.

The Pearson coefficient is a widely used measure of similarity. It allows
us to say when vertices are both similar or dissimilar compared with what we
would expect if connections in the network were formed at random.

7.12.3 O THER MEASURES OF STRUCTURAL EQUIVALENCE

There are many other possible measures of structural equivalence. For in-
stance, one could also normalize the number nij of common neighbors by di-
viding by (rather than subtracting) the expected value of k i k j /n. That would
give us a similarity of
nij ∑k Aik A jk
=n . (7.53)
k i k j /n ∑k Aik ∑k A jk

215
M EASURES AND METRICS

This quantity will be 1 if the number of common neighbors is exactly as ex-

pected on the basis of chance, greater than one if there are more common
neighbors than that, and less than one for dissimilar vertices with fewer com-
mon neighbors than we would expect by chance. It is never negative and has
the nice property that it is zero when the vertices in question have no common
neighbors. This measure could be looked upon as an alternative to the cosine
similarity: the two differ in that one has the product of the degrees √ k i k j in the
denominator while the other has the square root of the product k i k j . It has
been suggested that Eq. (7.53) may in some cases be a superior measure to the
cosine similarity because, by normalizing with respect to the expected number
of common neighbors rather than the maximum number, it allows us to eas-
ily identify statistically surprising coincidences between the neighborhoods of
vertices, which cosine similarity does not [195].
Another measure of structural equivalence is the so-called Euclidean dis-
tance,32 which is equal to the number of neighbors that differ between two
vertices. That is, it is the number of vertices that are neighbors of i but not of j,
or vice versa. Euclidean distance is really a dissimilarity measure, since it is
larger for vertices that differ more.
In terms of the adjacency matrix the Euclidean distance dij between two
vertices can be written
dij = ∑( Aik − A jk )2 . (7.54)
k

As with our other measures it is sometimes convenient to normalize the Eu-

clidean distance by dividing by its possible maximum value. The maximum
value of dij occurs when two vertices have no neighbors in common, in which
case the distance is equal to the sum of the degrees of the vertices: dij = k i + k j .
Dividing by this maximum value the normalized distance is

∑k ( Aik − A jk )2 ∑ ( Aik + A jk − 2Aik A jk ) nij

= k = 1−2 , (7.55)
ki + k j ki + k j ki + k j

where we have made use of the fact that A2ij = Aij because Aij is always zero or
one, and nij is again the number of neighbors that i and j have in common. To
within additive and multiplicative constants, this normalized Euclidean dis-
tance can thus be regarded as just another alternative normalization of the
number of common neighbors.

32
This is actually a bad name for it—it should be called Hamming distance, since it is essentially
the same as the Hamming distance of computer science and has nothing to do with Euclid.

216
 7.12 | S IMILARITY

7.12.4 R EGULAR EQUIVALENCE

The similarity measures discussed in the preceding sections are all measures of
structural equivalence, i.e., they are measures of the extent to which two ver-
tices share the same neighbors. The other main type of similarity considered
in social network analysis is regular equivalence. As described above, regu-
larly equivalent vertices are vertices that, while they do not necessarily share
neighbors, have neighbors who are themselves similar—see Fig. 7.9b again.
Quantitative measures of regular equivalence are less well developed than
measures of structural equivalence. In the 1970s social network analysts came
up with some rather complicated computer algorithms, such as the “REGE”
algorithm of White and Reitz [320, 327], that were intended to discover regular
equivalence in networks, but the operation of these algorithms is involved and k l
not easy to interpret. More recently, however, some simpler algebraic measures
have been developed that appear to work reasonably well. The basic idea [45, i j
162, 195] is to define a similarity score σij such that i and j have high similarity
if they have neighbors k and l that themselves have high similarity. For an
undirected network we can write this as Vertices i and j are consid-
ered similar (dashed line) if
σij = α ∑ Aik A jl σkl , (7.56) they have respective neigh-
kl
bors k and l that are them-
or in matrix terms σ = αAσA. Although it may not be immediately clear, selves similar.
this expression is a type of eigenvector equation, where the entire matrix σ of
similarities is the eigenvector. The parameter α is the eigenvalue (or more cor-
rectly, its inverse) and, as with the eigenvector centrality of Section 7.2, we are
See Section 11.1 for a dis-
normally interested in the leading eigenvalue, which can be found by standard cussion of computer algo-
methods. rithms for finding eigen-
This formula however has some problems. First, it doesn’t necessarily give vectors.
a high value for the “self-similarity” σii of a vertex to itself, which is counter-
intuitive. Presumably, all vertices are highly similar to themselves! As a conse-
quence of this, Eq. (7.56) also doesn’t necessarily give a high similarity score to
vertex pairs that have a lot of common neighbors, which in the light of our ex-
amination of structural equivalence in the preceding few sections we perhaps
feel it should. If we had high self-similarity scores for all vertices, on the other
hand, then Eq. (7.56) would automatically give high similarity also to vertices
with many common neighbors.
We can fix these problems by introducing an extra diagonal term in the
similarity thus:
σij = α ∑ Aik A jl σkl + δij , (7.57)
kl

217
M EASURES AND METRICS

or in matrix notation
σ = αAσA + I. (7.58)
However, while expressions like this have been proposed as similarity mea-
sures, they still suffer from some problems. Suppose we evaluate Eq. (7.58) by
repeated iteration, taking a starting value, for example, of σ (0) = 0 and using
it to compute σ (1) = αAσA + I, and then repeating the process many times
until σ converges. On the first few iterations we will get the following results:

σ (1) = I, (7.59a)
σ (2) = αA2 + I, (7.59b)
(3)
σ = α A + αA + I.
2 4 2
(7.59c)

The pattern is clear: in the limit of many iterations, we will get a sum over
even powers of the adjacency matrix. However, as discussed in Section 6.10,
k the elements of the rth power of the adjacency matrix count paths of length r
between vertices, and hence this measure of similarity is a weighted sum over
i j the numbers of paths of even length between pairs of vertices.
But why should we consider only paths of even length? Why not consider
paths of all lengths? These questions lead us to a better definition of regular
In the modified definition equivalence as follows: vertices i and j are similar if i has a neighbor k that is
of regular equivalence ver- itself similar to j.33 Again we assume that vertices are similar to themselves,
tex i is considered similar which we can represent with a diagonal δij term in the similarity, and our sim-
to vertex j (dashed line) if ilarity measure then looks like
it has a neighbor k that is it-
self similar to j. σij = α ∑ Aik σkj + δij , (7.60)
k

or
σ = αAσ + I, (7.61)
in matrix notation. Evaluating this expression by iterating again starting from
σ (0) = 0, we get

σ (1) = I, (7.62a)
(2)
σ = αA + I, (7.62b)
(3)
σ = α A + αA + I.
2 2
(7.62c)

33
This definition is not obviously symmetric with respect to i and j but, as we see, does in fact
give rise to an expression for the similarity that is symmetric.

218
 7.12 | S IMILARITY

In the limit of a large number of iterations this gives

∞
σ= ∑ (αA)m = (I − αA)−1 , (7.63)
m =0

which we could also have deduced directly by rearranging Eq. (7.61). Now
our similarity measure includes counts of paths at all lengths, not just even
paths. In fact, we can see now that this similarity measure could be defined
a completely different way, as a weighted count of all the paths between the
vertices i and j with paths of length r getting weight αr . So long as α < 1,
longer paths will get less weight than shorter ones, which seems sensible: in
effect we are saying that vertices are similar if they are connected either by a
few short paths or by very many long ones.
Equation (7.63) is reminiscent of the formula for the Katz centrality, Eq.
(7.10). We could call Eq. (7.63) the “Katz similarity” perhaps, although Katz
himself never discussed it. The Katz centrality of a vertex would then be sim-
ply the sum of the Katz similarities of that vertex to all others. Vertices that
are similar to many others would get high centrality, a concept that certainly
makes intuitive sense. As with the Katz centrality, the value of the parameter
α is undetermined—we are free to choose it as we see fit—but it must satisfy
α < 1/κ1 if the sum in Eq. (7.63) is to converge, where κ1 is the largest eigen-
value of the adjacency matrix.
In a sense, this regular equivalence measure can be seen as a generalization
of our structural equivalence measures in earlier sections. With those measures
we were counting the common neighbors of a pair of vertices, but the number
of common neighbors is also of course the number of paths of length two be-
tween the vertices. Our “Katz similarity” measure merely extends this concept
to counting paths of all lengths.
Some variations of this similarity measure are possible. As defined it tends
to give high similarity to vertices that have high degree, because if a vertex
has many neighbors it tends to increase the number of those neighbors that
are similar to any other given vertex and hence increases the total similarity
to that vertex. In some cases this might be desirable: maybe the person with
many friends should be considered more similar to others than the person with
few. However, in other cases it gives an unwanted bias in favor of high-degree
nodes. Who is to say that two hermits are not “similar” in an interesting sense?
If we wish, we can remove the bias in favor of high degree by dividing by
vertex degree thus:
α
σij = ∑ Aik σkj + δij , (7.64)
ki k

219
M EASURES AND METRICS

or in matrix notation σ = αD−1 Aσ + I, where, as previously, D is the diagonal

matrix with elements Dii = k i . This expression can be rearranged to read:34

σ = (I − αD−1 A)−1 = (D − αA)−1 D. (7.65)

Another useful variant is to consider cases where the last term in Eqs. (7.60)
or (7.64) is not simply diagonal, but includes off-diagonal terms too. Such
a generalization would allow us to specify explicitly that particular pairs of
vertices are similar, based on some other (probably non-network) information
that we have at our disposal. Going back to the example of CEOs at compa-
nies that we gave at the beginning of Section 7.12, we might, for example, want
to state explicitly that the CFOs and CIOs and so forth at different companies
are similar, and then our similarity measure would, we hope, correctly deduce
from the network structure that the CEOs are similar also. This kind of ap-
proach is particularly useful in the case of networks that consist of more than
one component, so that some pairs of vertices are not connected at all. If, for
instance, we have two separate components representing people in two differ-
ent companies, then there will be no paths of any length between individuals
in different companies, and hence a measure like (7.60) or (7.64) will never as-
sign a non-zero similarity to such individuals. If however, we explicitly insert
some similarities between members of the different companies, our measure
will then be able to generalize and extend those inputs to deduce similarities
between other members.
This idea of generalizing from a few given similarities arises in other con-
texts too. For example, in the fields of machine learning and information re-
trieval there is a considerable literature on how to generalize known similar-
ities between a subset of the objects in a collection of, say, text documents to
the rest of the collection, based on network data or other information.

7.13 H OMOPHILY AND ASSORTATIVE MIXING

Consider Fig. 7.10, which shows a friendship network of children at an Amer-
ican school, determined from a questionnaire of the type discussed in Sec-
tion 3.2.35 One very clear feature that emerges from the figure is the division of

34
It is interesting to note that when we expand this measure in powers of the adjacency matrix,
as we did in Eq. (7.63), the second-order (i.e., path-length two) term is the same as the structural
equivalence measure of Eq. (7.53), which perhaps lends further credence to both expressions as
natural measures of similarity.
35
The study used a “name generator”—students were asked to list the names of others they
considered to be their friends. This results in a directed network, but we have neglected the edge

220
 7.13 | H OMOPHILY AND ASSORTATIVE MIXING

Black
White
Other

Figure 7.10: Friendship network at a US high school. The vertices in this network represent 470 students at a US
high school (ages 14 to 18 years). The vertices are color coded by race as indicated in the key. Data from the National
Longitudinal Study of Adolescent Health [34, 314].

the network into two groups. It turns out that this division is principally along
lines of race. The different shades of the vertices in the picture correspond to
students of different race as denoted in the legend, and reveal that the school is
sharply divided between a group composed principally of black children and
a group composed principally of white.
This is not news to sociologists, who have long observed and discussed
such divisions [225]. Nor is the effect specific to race. People are found to
form friendships, acquaintances, business relations, and many other types of
tie based on all sorts of characteristics, including age, nationality, language, in-
come, educational level, and many others. Almost any social parameter you

directions in the figure. In our representation there is an undirected edge between vertices i and j
if either of the pair considers the other to be their friend (or both).

221
M EASURES AND METRICS

can imagine plays into people’s selection of their friends. People have, it ap-
pears, a strong tendency to associate with others whom they perceive as being
similar to themselves in some way. This tendency is called homophily or assor-
tative mixing.
More rarely, one also encounters disassortative mixing, the tendency for peo-
ple to associate with others who are unlike them. Probably the most widespread
and familiar example of disassortative mixing is mixing by gender in sexual
contact networks. The majority of sexual partnerships are between individu-
als of opposite sex, so they represent connections between people who differ
in their gender. Of course, same-sex partnerships do also occur, but they are a
much smaller fraction of the ties in the network.
Assortative (or disassortative) mixing is also seen in some nonsocial net-
works. Papers in a citation network, for instance, tend to cite other papers in
the same field more than they do papers in different fields. Web pages written
in a particular language tend to link to others in the same language.
In this section we look at how assortative mixing can be quantified. As-
sortative mixing by discrete characteristics such as race, gender, or nationality
is fundamentally different from mixing by a scalar characteristic like age or
income, so we treat the two cases separately.

7.13.1 A SSORTATIVE MIXING BY ENUMERATIVE CHARACTERISTICS

Suppose we have a network in which the vertices are classified according to
some characteristic that has a finite set of possible values. The values are
merely enumerative—they don’t fall in any particular order. For instance, the
vertices could represent people and be classified according to nationality, race,
or gender. Or they could be web pages classified by what language they are
written in, or biological species classified by habitat, or any of many other pos-
sibilities.
The network is assortative if a significant fraction of the edges in the net-
work run between vertices of the same type, and a simple way to quantify
assortativity would be to measure that fraction. However, this is not a very
good measure because, for instance, it is 1 if all vertices belong to the same
single type. This is a trivial sort of assortativity: all friends of a human be-
ing, for example, are also human beings,36 but this is not really an interesting
statement. What we would like instead is a measure that is large in non-trivial
cases but small in trivial ones.
A good measure turns out to be the following. We find the fraction of edges

36
Ignoring, for the purposes of argument, dogs, cats, imaginary friends, and so forth.

222
 7.13 | H OMOPHILY AND ASSORTATIVE MIXING

that run between vertices of the same type, and then we subtract from that fig-
ure the fraction of such edges we would expect to find if edges were positioned
at random without regard for vertex type. For the trivial case in which all ver-
tices are of a single type, for instance, 100% of edges run between vertices of
the same type, but this is also the expected figure, since there is nowhere else
for the edges to fall. The difference of the two numbers is then zero, telling us
that there is no non-trivial assortativity in this case. Only when the fraction of
edges between vertices of the same type is significantly greater than we would
expect on the basis of chance will our measure give a positive score.
In mathematical terms, let us denote by ci the class or type of vertex i, which
is an integer 1 . . . nc , with nc being the total number of classes. Then the total
number of edges that run between vertices of the same type is

∑ δ ( ci , c j ) = 1
2 ∑ Aij δ(ci , c j ), (7.66)
edges (i,j) ij

where δ(m, n) is the Kronecker delta and the factor of 12 accounts for the fact
that every vertex pair i, j is counted twice in the second sum.
Calculating the expected number of edges between vertices if edges are
placed at random takes a little more work. Consider a particular edge attached
to vertex i, which has degree k i . There are by definition 2m ends of edges in
the entire network, where m is as usual the total number of edges, and the
chances that the other end of our particular edge is one of the k j ends attached
to vertex j is thus k j /2m if connections are made purely at random.37 Counting
all k i edges attached to i, the total expected number of edges between vertices i
and j is then k i k j /2m, and the expected number of edges between all pairs of
vertices of the same type is

ki k j
1
2 ∑ 2m δ(ci , c j ), (7.67)
ij

where the factor of 12 , as before, prevents us from double-counting vertex pairs.

Taking the difference of (7.66) and (7.67) then gives us an expression for the
difference between the actual and expected number of edges in the network

37
Technically, we are making connections at random while preserving the vertex degrees. We
could in principle ignore vertex degrees and make connections truly at random, but in practice
this is found to give much poorer results.

223
M EASURES AND METRICS

that join vertices of like types:

 
ki k j ki k j
1
2 ∑ Aij δ(ci , c j ) − 1
2 ∑ 2m
δ ( ci , c j ) = 1
2 ∑ Aij −
2m
δ ( c i , c j ).
ij ij ij

(7.68)

Conventionally, one calculates not the number of such edges but the fraction,
which is given by this same expression divided by the number m of edges:
 
1 ki k j
2m ∑
Q= A ij − δ ( c i , c j ). (7.69)
ij
2m

This quantity Q is called the modularity [239,250] and is a measure of the extent
to which like is connected to like in a network. It is strictly less than 1, takes
positive values if there are more edges between vertices of the same type than
we would expect by chance, and negative ones if there are less.
For Fig. 7.10, for instance, where the types are the three ethnic classifica-
tions “black,” “white,” and “other,” we find a modularity value of Q = 0.305,
indicating (positive) assortative mixing by race in this particular network.38
Negative values of the modularity indicate disassortative mixing. We might
see a negative modularity, for example, in a network of sexual partnerships
where most partnerships were between individuals of opposite sex.
The quantity
ki k j
Bij = Aij − (7.70)
2m
in Eq. (7.69) appears in a number of situations in the study of networks. We will
encounter it, for instance, in Section 11.8 when we study community detection
in networks. In some contexts it is useful to consider Bij to be an element of a
matrix B, which itself is called the modularity matrix.
The modularity, Eq. (7.69), is always less than 1 but in general it does not
achieve the value Q = 1 even for a perfectly mixed network, one in which
every vertex is connected only to others of the same type. Depending on the
sizes of the groups and the degrees of vertices, the maximum value of Q can
be considerably less than 1. This is in some ways unsatisfactory: how is one to

38
An alternative measure of assortativity has been proposed by Gupta et al. [152]. That measure
however gives equal weight to each group of vertices, rather than to each edge as the modularity
does. With this measure if one had a million vertices of each of two types, which mixed with
one another entirely randomly, and ten more vertices of a third type that connected only among
themselves, one would end up with a score of about 0.5 [239], which appears to imply strong
assortativity when in fact almost all of the network mixes randomly. For most purposes therefore,
the measure of Eq. (7.69) gives results more in line with our intuitions.

224
 7.13 | H OMOPHILY AND ASSORTATIVE MIXING

know when one has strong assortative mixing and when one doesn’t? To rec-
tify the problem, we can normalize Q by dividing by its value for the perfectly
mixed network. With perfect mixing all edges fall between vertices of the same
type and hence δ(ci , c j ) = 1 whenever Aij = 1. This means that the first term
in the sum in Eq. (7.69) sums to 2m and the modularity for the perfectly mixed
network is  
1 ki k j
Qmax = 2m − ∑ δ ( ci , c j ) . (7.71)
2m ij
2m

Then the normalized value of the modularity is given by

Q ∑ij ( Aij − k i k j /2m)δ(ci , c j )

= . (7.72)
Qmax 2m − ∑ij (k i k j /2m)δ(ci , c j )

This quantity, sometimes called an assortativity coefficient, now takes a maxi-

mum value of 1 on a perfectly mixed network.
Although it can be a useful measure in some circumstances, however, Eq.
(7.72) is only rarely used. Most often, the modularity is used in its unnormal-
ized form, Eq. (7.69).
An alternative form for the modularity, which is sometimes useful in prac-
tical situations, can be derived in terms of the quantities

1
ers =
2m ∑ Aij δ(ci , r) δ(c j , s), (7.73)
ij

which is the fraction of edges that join vertices of type r to vertices of type s,
and
1
2m ∑
ar = k i δ ( c i , r ), (7.74)
i

which is the fraction of ends of edges attached to vertices of type r. Then,

noting that
δ ( c i , c j ) = ∑ δ ( c i , r ) δ ( c j , r ), (7.75)
r

we have, from Eq. (7.69)

 
1 ki k j
2m ∑ 2m ∑
Q= Aij − δ ( ci , r ) δ ( c j , r )
ij r
 
1 1 1
=∑ ∑ Aij δ(ci , r)δ(c j , r) − 2m ∑ ki δ(ci , r) 2m ∑ k j δ(c j , r )
r 2m ij i j
 
= ∑ err − ar . 2
(7.76)
r

225
M EASURES AND METRICS

This form can be useful, for instance, when we have network data in the form
of a list of edges and the types of the vertices at their ends, but no explicit data
on vertex degrees. In such a case ers and ar are relatively easy to calculate,
while Eq. (7.69) is quite awkward.

7.13.2 A SSORTATIVE MIXING BY SCALAR CHARACTERISTICS

We can also have homophily in a network according to scalar characteristics
like age or income. These are characteristics whose values come in a particular
order, so that it is possible say not only when two vertices are exactly the same
according to the characteristic but also when they are approximately the same.
For instance, while two people can certainly be of exactly the same age—born
on the same day even—they can also be approximately the same age—born
within a couple of years of one another, say—and people could (and in fact of-
ten do) choose who they associate with on the basis of such approximate ages.
There is no equivalent approximate similarity for the enumerative characteris-
tics of the previous section: there is no sense in which people from France and
Germany, say, are more nearly of the same nationality than people from France
and Spain.39
If network vertices with similar values of a scalar characteristic tend to be
connected together more often that those with different values then the net-
work is considered assortatively mixed according to that characteristic. If, for
example, people are friends with others around the same age as them, then the
network is assortatively mixed by age. Sometimes you may also hear it said
A sketch of stratified that the network is stratified by age, which means the same thing—one can
network in which most think of age as a one-dimensional scale or axis, with individuals of different
connections run between ages forming connected “strata” within the network.
vertices at or near the same Consider Fig. 7.11, which shows friendship data for the same set of US
“level” in the network, schoolchildren as Fig. 7.10 but now as a function of age. Each dot in the figure
with level along the ver-
corresponds to one pair of friends and the position of the dot along the two
tical axis in this case and
axes gives the ages of the friends, with ages measured by school grades.40 As
also denoted by the shades
of the vertices. the figure shows, there is substantial assortative mixing by age among the stu-
dents: many dots lie within the boxes close to the diagonal line that represent

39
Of course, one could make up some measure of national differences, based say on geographic
distance, but if the question we are asked is, “Are these two people of the same nationality?” then
under normal circumstances the only answers are “yes” and “no.” There is nothing in between.
40
In the US school system there are 12 grades of one year each and to begin grade g students
normally must be at least of age g + 5. Thus the 9th grade corresponds to children of age 14 and
15.

226
 7.13 | H OMOPHILY AND ASSORTATIVE MIXING

12

11
Age (grade)

10

9 10 11 12
Age (grade)

Figure 7.11: Ages of pairs of friends in high school. In this scatter plot each dot cor-
responds to one of the edges in Fig. 7.10, and its position along the horizontal and
vertical axes gives the ages of the two individuals at either end of that edge. The ages
are measured in terms of the grades of the students, which run from 9 to 12. In fact,
grades in the US school system don’t correspond precisely to age since students can
start or end their high-school careers early or late, and can repeat grades. (Each student
is positioned at random within the interval representing their grade, so as to spread the
points out on the plot. Note also that each friendship appears twice, above and below
the diagonal.)

friendships between students in the same grade. There is also, in this case, a
notable tendency for students to have more friends of a wider range of ages
as their age increases so there is a lower density of points in the top right box
than in the lower left one.
One could make a crude measure of assortative mixing by scalar charac-
teristics by adapting the ideas of the previous section. One could group the
vertices into bins according to the characteristic of interest (say age) and then
treat the bins as separate “types” of vertex in the sense of Section 7.13.1. For in-
stance, we might group people by age in ranges of one year or ten years. This
however misses much of the point about scalar characteristics, since it con-
siders vertices falling in the same bin to be of identical types when they may

227
M EASURES AND METRICS

be only approximately so, and vertices falling in different bins to be entirely

different when in fact they may be quite similar.
A better approach is to use a covariance measure as follows. Let xi be the
value for vertex i of the scalar quantity (age, income, etc.) that we are interested
in. Consider the pairs of values ( xi , x j ) for the vertices at the ends of each
edge (i, j) in the network and let us calculate their covariance over all edges as
follows. We define the mean μ of the value of xi at the end of an edge thus:

∑ij Aij xi ∑ k i xi 1
μ=
∑ij Aij
= i
∑i k i
=
2m ∑ k i xi . (7.77)
i

Note that this is not simply the mean value of xi averaged over all vertices. It
is an average over edges, and since a vertex with degree k i lies at the ends of k i
edges it appears k i times in the average (hence the factor of k i in the sum).
Then the covariance of xi and x j over edges is

∑ij Aij ( xi − μ)( x j − μ)

cov( xi , x j ) =
∑ij Aij
1
=
2m ∑ Aij (xi x j − μxi − μx j + μ2 )
ij
1
=
2m ∑ Aij xi x j − μ2
ij
1 1
=
2m ∑ Aij xi x j − (2m)2 ∑ ki k j xi x j
ij ij
 
1 ki k j
=
2m ∑ Aij −
2m
xi x j , (7.78)
ij

where we have made use of Eqs. (6.21) and (7.77). Note the strong similar-
ity between this expression and Eq. (7.69) for the modularity—only the delta
function δ(ci , c j ) in (7.69) has changed, being replaced by xi x j .
The covariance will be positive if, on balance, values xi , x j at either end of
an edge tend to be both large or both small and negative if they tend to vary in
opposite directions. In other words, the covariance will be positive when we
have assortative mixing and negative for disassortative mixing.
Just as with the modularity measure of Section 7.13.1, it is sometimes con-
venient to normalize the covariance so that it takes the value 1 in a perfectly
mixed network—one in which all edges fall between vertices with precisely
equal values of xi (although in most cases such an occurrence would be ex-
tremely unlikely in practice). Putting x j = xi in Eq. (7.78) gives a perfect mix-

228
 7.13 | H OMOPHILY AND ASSORTATIVE MIXING

ing value of
   
1 ki k j 2 1 ki k j
2m ∑ Aij −
2m
xi =
2m ∑ k i δij −
2m
xi x j , (7.79)
ij ij

and the normalized measure, sometimes called an assortativity coefficient, is the

ratio of the two:
∑ij ( Aij − k i k j /2m) xi x j
r= . (7.80)
∑ij (k i δij − k i k j /2m) xi x j
Although it may not be immediately obvious, this is in fact an example of a
(Pearson) correlation coefficient, having a covariance in its numerator and a
variance in the denominator. We encountered another example in a different
context in Section 7.12.2. The correlation coefficient varies in value between a
maximum of 1 for a perfectly assortative network and a minimum of −1 for a
perfectly disassortative one. A value of zero implies that the values of xi at the
ends of edges are uncorrelated.41
For the data of Fig. 7.11 the correlation coefficient is found to take a value
of r = 0.616, indicating that the student friendship network has significant
assortative mixing by age—students tend to be friends with others who have
ages close to theirs.
It would be possible in principle also to have assortative (or disassortative)
mixing according to vector characteristics, with vertices whose vectors have
similar values, as measured by some appropriate metric, being more (or less)
likely to be connected by an edge. One example of such mixing is the for-
mation of friendships between individuals according to their geographic lo-
cations, location being specified by a two-dimensional vector of, for example,
latitude/longitude coordinates. It is certainly the case that in general people
tend to be friends with others who live geographically close to them, so one
would expect mixing of this type to be assortative. Formal treatments of vec-
tor assortative mixing, however, have not been much pursued in the network
literature so far.

41
There could be non-linear correlations in such a network and we could still have r = 0; the
correlation coefficient detects only linear correlations. For instance, we could have vertices with
high and low values of xi connected predominantly to vertices with intermediate values. This is
neither assortative nor disassortative by the conventional definition and would give a small value
of r, but might nonetheless be of interest. Such non-linear correlations could be discovered by
examining a plot such as Fig. 7.11 or by using alternative measures of correlation such as informa-
tion theoretic measures. Thus it is perhaps wise not to rely solely on the value of r in investigating
assortative mixing.

229
M EASURES AND METRICS

7.13.3 A SSORTATIVE MIXING BY DEGREE

A special case of assortative mixing according to a scalar quantity, and one of
particular interest, is that of mixing by degree. In a network that shows assorta-
tive mixing by degree the high-degree vertices will be preferentially connected
to other high-degree vertices, and the low to low. In a social network, for exam-
ple, we have assortative mixing by degree if the gregarious people are friends
with other gregarious people and the hermits with other hermits. Conversely,
we could have disassortative mixing by degree, which would mean that the
gregarious people were hanging out with hermits and vice versa.
The reason this particular case is interesting is because, unlike age or in-
come, degree is itself a property of the network structure. Having one struc-
tural property (the degrees) dictate another (the positions of the edges) gives
rise to some interesting features in networks. In particular, in an assortative
network, where the high-degree nodes tend to stick together, one expects to
get a clump or core of such high-degree nodes in the network surrounded by
a less dense periphery of nodes with lower-degree. This core/periphery structure
is a common feature of social networks, many of which are found to be as-
sortatively mixed by degree. Figure 7.12a shows a small assortatively mixed
network in which the core/periphery structure is clearly visible.
On the other hand, if a network is disassortatively mixed by degree then
high-degree vertices tend to connected to low-degree ones, creating star-like
features in the network that are often readily visible. Figure 7.12b shows an
example of a small disassortative network. Disassortatively networks do not
usually have a core/periphery split but are instead more uniform.
Assortative mixing by degree can be measured in the same way as mixing
according to any other scalar quantity. We define a covariance of the type
described by Eq. (7.78), but with xi now equal to the degree k i :
 
1 ki k j
2m ∑
cov(k i , k j ) = A ij − ki k j , (7.81)
ij
2m

or if we wish we can normalize by the maximum value of the covariance to get

a correlation coefficient or assortativity coefficient:

∑ij ( Aij − k i k j /2m)k i k j

r= . (7.82)
∑ij (k i δij − k i k j /2m)k i k j
We give examples of the application of this formula to a number of networks
in Section 8.7.
One point to notice is that the evaluation of Eq. (7.81) or Eq. (7.82) requires
only the structure of the network and no other information (unlike the calcu-

230
 P ROBLEMS

(a) (b)

Figure 7.12: Assortative and disassortative networks. These two small networks are not real networks—they were
computer generated to display the phenomenon of assortativity by degree. (a) A network that is assortative by degree,
displaying the characteristic dense core of high-degree vertices surrounded by a periphery of lower-degree ones. (b) A
disassortative network, displaying the star-like structures characteristic of this case. Figure from Newman and Gir-
van [249]. Copyright 2003 Springer-Verlag Berlin Heidelberg. Reproduced with kind permission of Springer Science
and Business Media.

lations for other forms of assortative mixing). Once we know the adjacency
matrix (and hence the degrees) of all vertices we can calculate r. Perhaps for
this reason mixing by degree is one of the most frequently studied types of
assortative mixing.

P ROBLEMS
7.1 Consider a k-regular undirected network (i.e., a network in which every vertex has
degree k).

231
M EASURES AND METRICS

a) Show that the vector 1 = (1, 1, 1, . . .) is an eigenvector of the adjacency matrix

with eigenvalue k.
b) By making use of the fact that eigenvectors are orthogonal (or otherwise), show
that there is no other eigenvector that has all elements positive. The Perron–
Frobenius theorem says that the eigenvector with the largest eigenvalue always
has all elements non-negative (see footnote 2 on page 346), and hence the eigen-
vector 1 gives, by definition, the eigenvector centrality of our k-regular network
and the centralities are the same for every vertex.
c) Find the Katz centralities of all vertices in a k-regular network.
d) You should have found that, as with the eigenvector centrality, the Katz central-
ities of all vertices in the network are the same. Name a centrality measure that
could give different centrality values for different vertices in a regular network.

7.2 Suppose a directed network takes the form of a tree with all edges pointing inward
towards a central vertex:

What is the PageRank centrality of the central vertex in terms of the single parameter α
appearing in the definition of PageRank and the geodesic distances di from each vertex i
to the central vertex?

7.3 Consider an undirected tree of n vertices. A particular edge in the tree joins ver-
tices 1 and 2 and divides the tree into two disjoint regions of n1 and n2 vertices as
sketched here:

1 2
n1
n2

232
 P ROBLEMS

Show that the closeness centralities C1 and C2 of the two vertices, defined according to
Eq. (7.29), are related by
1 n1 1 n2
+ = + .
C1 n C2 n
7.4 Consider an undirected (connected) tree of n vertices. Suppose that a particular
vertex in the tree has degree k, so that its removal would divide the tree into k disjoint
regions, and suppose that the sizes of those regions are n1 . . . nk .
a) Show that the unnormalized betweenness centrality x of the vertex, as defined in
Eq. (7.36), is
k
x = n2 − ∑ n2m .
m =1
b) Hence, or otherwise, calculate the betweenness of the ith vertex from the end of a
“line graph” of n vertices, i.e., n vertices in a row like this:

7.5 Consider these three networks:

A B

a) Find a 3-core in the first network.

b) What is the reciprocity of the second network?
c) What is the cosine similarity of vertices A and B in the third network?

7.6 Among all pairs of vertices in a directed network that are connected by an edge or
edges, suppose that half are connected in only one direction and the rest are connected
in both directions. What is the reciprocity of the network?

7.7 In this network + and − indicate pairs of people who like each other or don’t,
respectively:

233
M EASURES AND METRICS

a) Is the network structurally balanced and why?

b) Is it clusterable and, if so, what are the clusters?

7.8 In a survey of couples in the US city of San Francisco, Catania et al. [65] recorded,
among other things, the ethnicity of their interviewees and calculated the fraction of
couples whose members were from each possible pairing of ethnic groups. The frac-
tions were as follows:
Women
Black Hispanic White Other Total
Black 0.258 0.016 0.035 0.013 0.323
Men Hispanic 0.012 0.157 0.058 0.019 0.247
White 0.013 0.023 0.306 0.035 0.377
Other 0.005 0.007 0.024 0.016 0.053
Total 0.289 0.204 0.423 0.084
Assuming the couples interviewed to be a representative sample of the edges in the
undirected network of relationships for the community studied, and treating the ver-
tices as being of four types—black, Hispanic, white, and other—calculate the numbers
err and ar that appear in Eq. (7.76) for each type. Hence calculate the modularity of the
network with respect to ethnicity.

234