Community structure in social and

biological networks
M. Girvan*†‡ and M. E. J. Newman*§
*Santa Fe Institute, 1399 Hyde Park Road, Santa Fe, NM 87501; †Department of Physics, Cornell University, Clark Hall, Ithaca, NY 14853-2501; and
§Department of Physics, University of Michigan, Ann Arbor, MI 48109-1120

Edited by Lawrence A. Shepp, Rutgers, State University of New Jersey–New Brunswick, Piscataway, NJ, and approved April 6, 2002 (received for review
December 6, 2001)

A number of recent studies have focused on the statistical prop- In this article, we consider another property, which, as we will
erties of networked systems such as social networks and the show, appears to be common to many networks, the property of
Worldwide Web. Researchers have concentrated particularly on a community structure. (This property is also sometimes called
few properties that seem to be common to many networks: the clustering, but we refrain from this usage to avoid confusion with
small-world property, power-law degree distributions, and net- the other meaning of the word clustering introduced in the
work transitivity. In this article, we highlight another property that preceding paragraph.) Consider for a moment the case of social
is found in many networks, the property of community structure, networks—networks of friendships or other acquaintances be-
in which network nodes are joined together in tightly knit groups, tween individuals. It is a matter of common experience that such
between which there are only looser connections. We propose a networks seem to have communities in them: subsets of vertices
method for detecting such communities, built around the idea of within which vertex–vertex connections are dense, but between
using centrality indices to find community boundaries. We test our which connections are less dense. A figurative sketch of a
method on computer-generated and real-world graphs whose network with such a community structure is shown in Fig. 1.
community structure is already known and find that the method (Certainly it is possible that the communities themselves also
detects this known structure with high sensitivity and reliability. join together to form metacommunities, and that those meta-
We also apply the method to two networks whose community communities are themselves joined together, and so on in a
structure is not well known—a collaboration network and a food hierarchical fashion. This idea is discussed further in the next
web—and find that it detects significant and informative commu- section.) The ability to detect community structure in a network
nity divisions in both cases. could clearly have practical applications. Communities in a social
network might represent real social groupings, perhaps by
interest or background; communities in a citation network (19)
M any systems take the form of networks, sets of nodes or
vertices joined together in pairs by links or edges (1).
Examples include social networks (2–4) such as acquaintance
might represent related papers on a single topic; communities in
a metabolic network might represent cycles and other functional
networks (5) and collaboration networks (6), technological groupings; communities on the web might represent pages
networks such as the Internet (7), the Worldwide Web (8, 9), and on related topics. Being able to identify these communities
power grids (4, 5), and biological networks such as neural could help us to understand and exploit these networks more
networks (4), food webs (10), and metabolic networks (11, 12). effectively.
Recent research on networks among mathematicians and phys- In this article we propose a method for detecting community

MATHEMATICS
icists has focused on a number of distinctive statistical properties structure and apply it to the study of a number of different social

APPLIED
that most networks seem to share. One such property is the and biological networks. As we will show, when applied to
‘‘small world effect,’’ which is the name given to the finding that networks for which the community structure is already known
the average distance between vertices in a network is short (13, from other studies, our method appears to give excellent agree-
14), usually scaling logarithmically with the total number n of ment with the expected results. When applied to networks for
vertices. Another is the right-skewed degree distributions that which we do not have other information about communities, it
many networks possess (8, 9, 15–17). The degree of a vertex in gives promising results that may help us understand better the
a network is the number of other vertices to which it is connected, interplay between network structure and function.
and one finds that there are typically many vertices in a network
with low degree and a small number with high degree, the precise Detecting Community Structure
distribution often following a power-law or exponential form In this section we review existing methods for detecting com-
(1, 5, 15). munity structure and discuss the ways in which these approaches
A third property that many networks have in common is may fail, before describing our own method, which avoids some
clustering, or network transitivity, which is the property that two of the shortcomings of the traditional techniques.
vertices that are both neighbors of the same third vertex have a
heightened probability of also being neighbors of one another. Traditional Methods. The traditional method for detecting com-
In the language of social networks, two of your friends will have munity structure in networks such as that depicted in Fig. 1 is
a greater probability of knowing one another than will two hierarchical clustering. One first calculates a weight Wij for every
people chosen at random from the population, on account of pair i,j of vertices in the network, which represents in some sense
their common acquaintance with you. This effect is quantified by how closely connected the vertices are. (We give some examples
the clustering coefficient C (4, 18), defined by of possible such weights below.) Then one takes the n vertices in
the network, with no edges between them, and adds edges
3⫻ (number of triangles on the graph) between pairs one by one in order of their weights, starting with
C⫽ . [1]
(number of connected triples of vertices) the pair with the strongest weight and progressing to the weakest.
As edges are added, the resulting graph shows a nested set of
This number is precisely the probability that two of one’s friends
are friends themselves. It is 1 on a fully connected graph
(everyone knows everyone else) and has typical values in the This paper was submitted directly (Track II) to the PNAS office.
range of 0.1 to 0.5 in many real-world networks. ‡To whom reprint requests should be addressed. E-mail: [email protected].

www.pnas.org兾cgi兾doi兾10.1073兾pnas.122653799 PNAS 兩 June 11, 2002 兩 vol. 99 兩 no. 12 兩 7821–7826

 computed quickly by using polynomial-time ‘‘max-flow’’ algo-
rithms such as the augmenting path algorithm (22).
Another possible way to define weights between vertices is to
count the total number of paths that run between them (all
paths, not just those that are node- or edge-independent).
However, because the number of paths between any two vertices
is infinite (unless it is zero), one typically weights paths of length
ᐉ by a factor ␣ᐉ with ␣ small, so that the weighted count of the
number of paths converges (23). Thus long paths contribute
exponentially less weight than those that are short. If A is the
adjacency matrix of the network, such that Aij is 1 if there is an
edge between vertices i and j and 0 otherwise, then the weights
in this definition are given by the elements of the matrix

冘
⬁

W⫽ 共␣A兲ᐉ ⫽ 关I ⫺ ␣A兴 ⫺ 1. [2]

ᐉ⫽0

For the sum to converge, we must choose ␣ smaller than the

reciprocal of the largest eigenvalue of A.
Both of these definitions of the weights give reasonable results
Fig. 1. A schematic representation of a network with community structure.
for community structure in some cases. In other cases they are
In this network there are three communities of densely connected vertices
(circles with solid lines), with a much lower density of connections (gray lines)
less successful. In particular, both have a tendency to separate
between them. single peripheral vertices from the communities to which they
should rightly belong. If a vertex is, for example, connected to the
rest of a network by only a single edge then, to the extent that
it belongs to any community, it should clearly be considered to
increasingly large components (connected subsets of vertices),
belong to the community at the other end of that edge. Unfor-
which are taken to be the communities. Because the components
tunately, both the numbers of independent paths and the
are properly nested, they all can be represented by using a tree weighted path counts for such vertices are small and hence single
of the type shown in Fig. 2, in which the lowest level at which two nodes often remain isolated from the network when the com-
vertices are connected represents the strength of the edge that munities are constructed. This and other pathologies, along with
resulted in their first becoming members of the same commu- poor results from these methods in some networks where the
nity. A ‘‘slice’’ through this tree at any level gives the commu- community structure is well known from other studies, make the
nities that existed just before an edge of the corresponding hierarchical clustering method, although useful, far from perfect.
weight was added. Trees of this type are sometimes called
dendrograms in the sociological literature. Edge ‘‘Betweenness’’ and Community Structure. To sidestep the
Many different weights have been proposed for use with shortcomings of the hierarchical clustering method, we here
hierarchical clustering algorithms. One possible definition of the propose an alternative approach to the detection of communi-
weight is the number of node-independent paths between ver- ties. Instead of trying to construct a measure that tells us which
tices. Two paths that connect the same pair of vertices are said edges are most central to communities, we focus instead on those
to be node-independent if they share none of the same vertices edges that are least central, the edges that are most ‘‘between’’
other than their initial and final vertices. One can similarly also communities. Rather than constructing communities by adding
count edge-independent paths. It is known (20) that the number the strongest edges to an initially empty vertex set, we construct
of node-independent (edge-independent) paths between two them by progressively removing edges from the original graph.
vertices i and j in a graph is equal to the minimum number of Vertex betweenness has been studied in the past as a measure
vertices (edges) that must be removed from the graph to of the centrality and influence of nodes in networks. First
disconnect i and j from one another. Thus these numbers are in proposed by Freeman (24), the betweenness centrality of a vertex
i is defined as the number of shortest paths between pairs of
a sense a measure of the robustness of the network to deletion
other vertices that run through i. It is a measure of the influence
of nodes (edges) (21). Numbers of independent paths can be
of a node over the flow of information between other nodes,
especially in cases where information flow over a network
primarily follows the shortest available path.
To find which edges in a network are most between other pairs
of vertices, we generalize Freeman’s betweenness centrality to
edges and define the edge betweenness of an edge as the number
of shortest paths between pairs of vertices that run along it. If
there is more than one shortest path between a pair of vertices,
each path is given equal weight such that the total weight of all
of the paths is unity. If a network contains communities or
groups that are only loosely connected by a few intergroup edges,
then all shortest paths between different communities must go
along one of these few edges. Thus, the edges connecting
communities will have high edge betweenness. By removing
Fig. 2. An example of a small hierarchical clustering tree. The circles at the these edges, we separate groups from one another and so reveal
bottom represent the vertices in the network, and the tree shows the order in the underlying community structure of the graph.
which they join together to form communities for a given definition of the The algorithm we propose for identifying communities is
weight Wij of connections between vertex pairs. simply stated as follows:

7822 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.122653799 Girvan and Newman

 1. Calculate the betweenness for all edges in the network.
2. Remove the edge with the highest betweenness.
3. Recalculate betweennesses for all edges affected by the
removal.
4. Repeat from step 2 until no edges remain.
As a practical matter, we calculate the betweennesses by using
the fast algorithm of Newman (25), which calculates betweenness
for all m edges in a graph of n vertices in time O(mn). Because
this calculation has to be repeated once for the removal of each
edge, the entire algorithm runs in worst-case time O(m2n).
However, after the removal of each edge, we only have to
recalculate the betweennesses of those edges that were affected
by the removal, which is at most only those in the same
component as the removed edge. This means that running time
may be better than worst-case for networks with strong com-
munity structure (those that rapidly break up into separate
components after the first few iterations of the algorithm).
To try to reduce the running time of the algorithm further, one
might be tempted to calculate the betweennesses of all edges only
once and then remove them in order of decreasing betweenness.
We find, however, that this strategy does not work well, because Fig. 3. The fraction of vertices correctly classified in computer-generated
if two communities are connected by more than one edge, then graphs of the type described in the text, as the average number of intercom-
there is no guarantee that all of those edges will have high munity edges per vertex is varied. The circles are results for the method
betweenness—we only know that at least one of them will. By presented in this article; the squares are for a standard hierarchical clustering
recalculating betweennesses after the removal of each edge we calculation based on numbers of edge-independent paths between vertices.
ensure that at least one of the remaining edges between two Each point is an average over 100 realizations of the graphs. Lines between
points are included solely as a guide to the eye.
communities will always have a high value.

Tests of the Method

club study of Zachary (26). In this study, Zachary observed 34
In this section we present a number of tests of our algorithm on
members of a karate club over a period of 2 years. During the
computer-generated graphs and on real-world networks for
course of the study, a disagreement developed between the
which the community structure is already known. In each case we
administrator of the club and the club’s instructor, which ulti-
find that our algorithm reliably detects the known structure.
mately resulted in the instructor’s leaving and starting a new
club, taking about a half of the original club’s members with him.
Computer-Generated Graphs. To test the performance of our
Zachary constructed a network of friendships between mem-
algorithm we have applied it to a large set of artificial, computer-
generated graphs similar to those depicted in Fig. 1. Each graph bers of the club, using a variety of measures to estimate the
was constructed with 128 vertices divided into four communities strength of ties between individuals. Here we use a simple
of 32 vertices each. Edges were placed between vertex pairs unweighted version of his network and apply our algorithm to it
in an attempt to identify the factions involved in the split of club.

MATHEMATICS
independently at random, with probability Pin for vertices be-
Fig. 4a shows the network, with the instructor and the admin-

APPLIED
longing to the same community and Pout for vertices in different
communities, with Pout ⬍ Pin. The probabilities were chosen so istrator represented by nodes 1 and 34, respectively. Fig. 4b shows
as to keep the average degree z of a vertex equal to 16. This the hierarchical tree of communities produced by our method.
produces graphs that have known community structure, but The most fundamental split in the network is the first one at the
which are essentially random in other respects. Feeding these top of the tree, which divides the network into two groups of
graphs into our algorithm, we measured the fraction of vertices roughly equal size. This split corresponds almost perfectly with
that were classified by the algorithm into their correct commu- the actual division of the club members after the break-up, as
nities, as a function of the average number of intercommunity revealed by which club they attended afterward. Only one node,
edges per vertex. The results are shown in Fig. 3 (circles). As Fig. node 3, is classified incorrectly. In other words, the application
3 shows, the algorithm performs nearly perfectly when zout ⬍ 6, of our algorithm to the empirically observed network of friend-
classifying 90% or more of the vertices correctly. Only for zout ⱖ ships is a good predictor of the subsequent social evolution of the
6 does the fraction correctly classified start to fall off substan- group.
tially. In other words, the algorithm performs very well almost For comparison we also have performed a traditional hierar-
to the point at which each vertex has as many intercommunity as chical clustering based on edge-independent paths for the karate
intracommunity connections. club network; the resulting tree is shown in Fig. 4c. As Fig. 4c
For comparison we also show in Fig. 3 (squares) the fraction shows, this method correctly identifies the core vertex sets
of vertices classified correctly by a standard hierarchical clus- {1,2,3} and {33,34} of the two communities, but otherwise there
tering calculation based on independent path counts computed appears to be little correlation with the actual split of the club,
by using max-flow. As Fig. 3 shows, the performance of this indicating once again that our method is significantly more
method is far inferior to that of our method. accurate and sensitive than the standard method.

Zachary’s Karate Club Study. Although computer-generated net- College Football. As a further test of our algorithm, we turn to the
works provide a reproducible and well controlled test bed for our world of United States college football. (Football here means
community-structure algorithm, it is clearly desirable to test the American football, not soccer.) The network we look at is a
algorithm on data from real-world networks as well. To this end, representation of the schedule of Division I games for the 2000
we have selected two datasets representing real-world networks season: vertices in the graph represent teams (identified by their
for which the community structure is already known from other college names) and edges represent regular-season games be-
sources. The first of these is drawn from the well known karate tween the two teams they connect. What makes this network

Girvan and Newman PNAS 兩 June 11, 2002 兩 vol. 99 兩 no. 12 兩 7823
Fig. 4. (a) The friendship network from Zachary’s karate club study (26) as
described in the text. Nodes associated with the club administrator’s faction
are drawn as circles, those associated with the instructor’s faction are drawn
as squares. (b) Hierarchical tree showing the complete community structure
for the network calculated by using the algorithm presented in this article. The
initial split of the network into two groups is in agreement with the actual
factions observed by Zachary, with the exception that node 3 is misclassified.
(c) Hierarchical tree calculated by using edge-independent path counts, which
fails to extract the known community structure of the network.

interesting is that it incorporates a known community structure.

The teams are divided into conferences containing around 8–12
teams each. Games are more frequent between members of the Fig. 5. Hierarchical tree for the network reflecting the schedule of regular-
same conference than between members of different confer- season Division I college football games for year 2000. Nodes in the network
ences, with teams playing an average of about seven intracon- represent teams, and edges represent games between teams. Our algorithm
ference games and four interconference games in the 2000 identifies nearly all of the conference structure in the network.
season. Interconference play is not uniformly distributed; teams
that are geographically close to one another but belong to
correspond to nuances in the scheduling of games. For example,
different conferences are more likely to play one another than the Sunbelt Conference is broken into two pieces and grouped
teams separated by large geographic distances. with members of the Western Athletic Conference. This happens
Applying our algorithm to this network, we find that it because the Sunbelt teams played nearly as many games against
identifies the conference structure with a high degree of success Western Athletic teams as they did against teams in their own
(Fig. 5). Almost all teams are correctly grouped with the other conference. They also played quite a large fraction of their
teams in their conference. There are a few independent teams interconference games against Mid-American teams. Naturally,
that do not belong to any conference—these tend to be grouped our algorithm fails in cases like this where the network structure
with the conference with which they are most closely associated. genuinely does not correspond to the conference structure. In all
The few cases in which the algorithm seems to fail actually other respects, however, it performs remarkably well.

7824 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.122653799 Girvan and Newman

 spond roughly with the split between economics and traffic. The
community represented by circles is comprised of a group of
scientists working on mathematical models in ecology and forms
a fairly cohesive structure, as evidenced by the fact that the
algorithm does not break it into smaller components to any
significant extent. The largest community, indicated by squares
of various shades, represents a group working primarily in
statistical physics and is subdivided into well defined smaller
groups, which are denoted by the different shadings. In this case,
each subcommunity seems to revolve around the research in-
terests of one dominant member. The community represented by
triangles is a group working primarily on the structure of RNA.
It, too, can be divided further into smaller subcommunities,
centered once again around the interests of leading members.
Our algorithm thus seems to find two types of communities:
scientists grouped together by similarity either of research topic
or methodology. It is not surprising to see communities built
around research topics; we expect scientists to collaborate
primarily with others with whom their research focus is closely
aligned. The formation of communities around methodologies is
more interesting and may be the mark of truly interdisciplinary
work. For example, the grouping of those working on economics
with those working on traffic models may seem surprising, until
one realizes that the technical approaches these scientists have
taken are quite similar. As a result of these kinds of similarities,
the network contains ties between researchers from traditionally
disparate fields. We conjecture that this feature may be peculiar
to interdisciplinary centers like the Santa Fe Institute.
Fig. 6. The largest component of the Santa Fe Institute collaboration
network, with the primary divisions detected by our algorithm indicated by Food Web. We have also applied our algorithm to a food web of
different vertex shapes.
marine organisms living in the Chesapeake Bay, a large estuary
on the east coast of the United States. This network was
Applications originally compiled by Baird and Ulanowicz (27) and contains 33
vertices representing the ecosystem’s most prominent taxa. Most
In the previous section we tested our algorithm on a number of taxa are represented at the species or genus level, although some
networks for which the community structure was known before- vertices represent larger groups of related species. Edges be-
hand. The results indicate that our algorithm is a sensitive and tween taxa indicate trophic relationships—one taxon feeding on
accurate method for extracting community structure from both another. Although relationships of this kind are inherently
real and artificial networks. In this section, we apply our method directed, we here ignore direction and consider the network to
to two more networks for which the structure is not known and

MATHEMATICS
be undirected.
show that in these cases it can help us to understand the make-up

APPLIED
Applying our algorithm to this network, we find two well
of otherwise complex and tangled datasets. Our first example is defined communities of roughly equal size, plus a small number
a collaboration network of scientists; our second is a food web of vertices that belong to neither community (see Fig. 7). As Fig.
of marine organisms in the Chesapeake Bay. 7 shows, the split between the two large communities corre-
sponds quite closely with the division between pelagic organisms
Collaboration Network. We have applied our community-finding (those that dwell principally near the surface or in the middle
method to a collaboration network of scientists at the Santa Fe depths of the bay) and benthic organisms (those that dwell near
Institute, an interdisciplinary research center in Santa Fe, New the bottom). Interestingly, the algorithm includes within each
Mexico (and current academic home to both authors of this group organisms from a variety of different trophic levels. This
article). The 271 vertices in this network represent scientists in finding contrasts with other techniques that have been used to
residence at the Santa Fe Institute during any part of calendar analyze food webs (28), which tend to cluster taxa according to
year 1999 or 2000 and their collaborators. An edge is drawn trophic level rather than habitat. Our results seem to imply that
between a pair of scientists if they coauthored one or more pelagic and benthic organisms in the Chesapeake Bay can be
articles during the same time period. The network includes all separated into reasonably self-contained ecological subsystems.
journal and book publications by the scientists involved, along The separation is not perfect: a small number of benthic
with all papers that appeared in the institute’s technical reports organisms find their way into the pelagic community, presumably
series. On average, each scientist coauthored articles with ap- indicating that these species play a substantial role in the food
proximately five others. chains of their surface-dwelling colleagues. This finding suggests
In Fig. 6 we illustrate the results from the application of our that the simple traditional division of taxa into pelagic or benthic
algorithm to the largest component of the collaboration graph may not be an ideal classification in this case.
(which consists of 118 scientists). Vertices are drawn as different We also have applied our method to a number of other food
shapes according to the primary divisions detected. We find that webs. Interestingly, although some of these show clear commu-
the algorithm splits the network into a few strong communities, nity structure similar to that of Fig. 7, some others do not. This
with the divisions running principally along disciplinary lines. could be because some ecosystems are genuinely not composed
The community indicated by diamonds is the least well defined of separate communities, but it also could be because many food
and represents a group of scientists using agent-based models to webs, unlike other networks, are dense, i.e., the number of edges
study problems in economics and traffic flow. The algorithm scales as the square of the number of vertices rather than scaling
further divides this group into smaller components that corre- linearly (29). Our algorithm was designed with sparse networks

Girvan and Newman PNAS 兩 June 11, 2002 兩 vol. 99 兩 no. 12 兩 7825
 known community structure with a high degree of success. We
have also tested it on two real-world networks with well docu-
mented structure and find the results to be in excellent agree-
ment with expectations. In addition, we have given two examples
of applications of the algorithm to networks whose structure was
not previously well documented and find that in both cases it
extracts clear communities that appear to correspond to plau-
sible and informative divisions of the network nodes.
A number of extensions or improvements of our method may
be possible. First, we hope to generalize the method to handle
both weighted and directed graphs. Second, we hope that it may
be possible to improve the speed of the algorithm. At present, the
algorithm runs in time O(n3) on sparse graphs, where n is the
number of vertices in the network. This makes it impractical for
very large graphs. Detecting communities in, for instance, the
large collaboration networks (6) or subsets of the web graph (9)
that have been studied recently, would be entirely unfeasible.
Perhaps, however, the basic principles of our approach—
focusing on the boundaries of communities rather than their
cores, and making use of edge betweenness—can be incorpo-
rated into a modified method that scales more favorably with
network size.
Fig. 7. Hierarchical tree for the Chesapeake Bay food web described in the
text. We hope that the ideas and methods presented here will prove
useful in the analysis of many other types of networks. Possible
further applications range from the determination of functional
in mind, and it is possible that it may not perform as well on clusters within neural networks to analysis of communities on the
dense networks. Worldwide Web, as well as others not yet thought of. We hope
to see such applications in the future.
Conclusions
In this article we have investigated community structure in We thank Jennifer Dunne, Neo Martinez, Matthew Salganik, Steve
networks of various kinds, introducing a method for detecting Strogatz, and Doug White for useful conversations, and Jennifer Dunne,
such structure. Unlike previous methods that focus on finding Sarah Knutson, Matthew Salganik, and Doug White for help with
the strongly connected cores of communities, our approach compiling the data for the food web, collaboration, college football, and
works by using information about edge betweenness to detect karate club networks, respectively. This work was funded in part by
community peripheries. We have tested our method on National Science Foundation Grants DMS-0109086, DGE-9870631, and
computer-generated graphs and have shown that it detects the PHY-9910217.

1. Strogatz, S. H. (2001) Nature (London) 410, 268–276. 14. Milgram, S. (1967) Psychol. Today 2, 60–67.
2. Wasserman, S. & Faust, K. (1994) Social Network Analysis (Cambridge Univ. 15. Barabási, A.-L. & Albert, R. (1999) Science 286, 509–512.
Press, Cambridge, U.K.). 16. Krapivsky, P. L., Redner, S. & Leyvraz, F. (2000) Phys. Rev. Lett. 85, 4629–4632.
3. Scott, J. (2000) Social Network Analysis: A Handbook (Sage, London), 2nd Ed. 17. Dorogovtsev, S. N., Mendes, J. F. F. & Samukhin, A. N. (2000) Phys. Rev. Lett.
4. Watts, D. J. & Strogatz, S. H. (1998) Nature (London) 393, 440–442. 85, 4633–4636.
5. Amaral, L. A. N., Scala, A., Barthélémy, M. & Stanley, H. E. (2000) Proc. Natl. 18. Newman, M. E. J., Strogatz, S. H. & Watts, D. J. (2001) Phys. Rev. E 64, 026118.
Acad. Sci. USA 97, 11149–11152. 19. Redner, S. (1998) Eur. Phys. J. B 4, 131–134.
6. Newman, M. E. J. (2001) Proc. Natl. Acad. Sci. USA 98, 404–409. 20. Menger, K. (1927) Fundamenta Mathematicae 10, 96–115.
7. Faloutsos, M., Faloutsos, P. & Faloutsos, C. (1999) Comput. Commun. Rev. 29, 21. White, D. R. & Harary, F. (2001) Sociol. Methodol. 31, 305–359.
251–262. 22. Ahuja, R. K., Magnanti, T. L. & Orlin, J. B. (1993) Network Flows: Theory,
8. Albert, R., Jeong, H. & Barabási, A.-L. (1999) Nature (London) 401, 130–131. Algorithms, and Applications (Prentice–Hall, Upper Saddle River, NJ).
9. Broder, A., Kumar, R., Maghoul, F., Raghavan, P., Rajagopalan, S., Stata, R., 23. Katz, L. (1953) Psychometrika 18, 39–43.
Tomkins, A. & Wiener, J. (2000) Comput. Networks 33, 309–320. 24. Freeman, L. (1977) Sociometry 40, 35–41.
10. Williams, R. J. & Martinez, N. D. (2000) Nature (London) 404, 180–183. 25. Newman, M. E. J. (2001) Phys. Rev. E 64, 016131.
11. Jeong, H., Tombor, B., Albert, R., Oltvai, Z. N. & Barabási, A.-L. (2000) Nature 26. Zachary, W. W. (1977) J. Anthropol. Res. 33, 452–473.
(London) 407, 651–654. 27. Baird, D. & Ulanowicz, R. E. (1989) Ecol. Monogr. 59, 329–364.
12. Fell, D. A. & Wagner, A. (2000) Nat. Biotechnol. 18, 1121–1122. 28. Yodzis, P. & Winemiller, K. O. (1999) Oikos 87, 327–340.
13. Pool, I. de S. & Kochen, M. (1978) Social Networks 1, 1–48. 29. Martinez, N. D. (1992) Am. Nat. 139, 1208–1218.

7826 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.122653799 Girvan and Newman