An Introduction to Graph Theory in Complex

Systems Studies

Table of contents
Why use a graph-theoretic representation?..................................................1
Representing a problem as a graph..............................................................2
Obtaining metrics through graph-theoretic analysis....................................3
Case study: metrics in a genetic regulatory network...............................4
Summary: steps in representing a problem as a graph.................................5
Tools for analysing graphs...........................................................................5
Programs..................................................................................................6
Libraries..................................................................................................8
Bibliography................................................................................................8
Glossary.......................................................................................................8

Note on terminology
Graph-theoretic or complex-systems terminology is identified by the use
of SMALL CAPS in the text. While brief definitions of these terms are
available in the glossary section, this introduction is intended to be used in
conjunction with the online reference material at:
http://130.102.66.173/wiki/index.php/Main_Page
This reference material provides definitions for common graph-theoretic
terms, and explanations of many complex systems metrics and techniques
for analysing graph-based representations.

Why use a graph-theoretic representation?

Graphs are a fundamental construct in complex systems research, and the use of graph
theoretic algorithms and metrics to extract useful information from a problem is a primary
method of analysis for complex systems. As with all problem representations, a graph-
based representation is used to provide a particular perspective on a problem. Whether or
not a graph-theoretic representation is useful for a particular problem depends on whether
the interesting features of the problem are likely to be highlighted by the kind of insights
that a graph-based representation may provide. Generally, graph-theoretic representations
of a problem are biased towards highlighting aspects of a problem that are either
structural (e.g., the connectivity of components or processes in the problem), or that
depend on structural properties (e.g., structure-dependant dynamics). The most common
types of problem in which structural aspects are important are problems involving
multiple interacting agents. In complex systems disciplines, for example, graph
representations are used because of their ability to analyse the patterns of interactions in
systems, as represented by the graph structure, on both a local and a global level.
Global graph metrics are used to provide information about general system properties due
to the structure of the whole graph. For example, a small RADIUS or DIAMETER indicates that
information will propagate very rapidly around the graph, while a SCALE-FREE ARCHITECTURE
indicates that the graph is likely to be robust to random damage. In contrast, local metrics
in a graph quantify characteristics of particular components of the graph, such as its
nodes, edges or subgraphs. For example, an edge acting as a BRIDGE is important because
its removal will cause the graph to become disconnected. Likewise, a subgraph that forms
a CLIQUE is considered interesting because the high interconnectedness of the nodes
suggests that the subgraph forms an important structure.
Some examples of systems amenable to graph-theoretic analysis are genetic regulatory
networks and social networks. In genetic regulatory networks, the global structure of
networks can provide insights into the potential role of self-organisation processes in the
development of regulatory systems. In social networks, an important focus is on the
analysis of local structures to determine important actors (nodes) and groups in the
network.
As these examples suggest, it is important to know what kind of global and local
structural properties may be expected in the problem at hand. This knowledge can then be
used to decide on the techniques to be used in analysing the problem.

Representing a problem as a graph

As a mathematical construct, a graph consists of two types of elements: nodes and edges.
In translating a problem to a graph-based representation, the first step is to decide how
problem elements will translate to these distinct graph elements. As a general guideline,
nodes are used to represent entities in a problem – genes, people, cities, businesses – and
edges are used to represent relationships between the entities – 'regulates', 'knows',
'exports to', 'sells to'. If these relationships are directional (e.g., the fact that A sells to B
doesn't imply that B sells to A), then the result will be a DIRECTED GRAPH; if not, then an
UNDIRECTED GRAPH will result. This distinction is important, as some graph-theoretic
measurements will treat directed and undirected graphs differently. Unless specifically
stated, undirected edges are the assumed norm.
In determining the mapping from problem elements to graph elements, it is sometimes
necessary to have multiple types of nodes or edges. While many problems can be
adequately represented without multiple node or edge types, some disciplines, such as
social network analysis, make significant use of 2-mode graphs (i.e., graphs having two
different node types). One example of a 2-mode graph is researchers and their papers, in
which both researchers and papers are represented as nodes, and a connection between a
researcher and a paper means that the researcher is an author of the paper. Having
multiple node or edge types may complicate analysis, as the techniques and metrics are
more well developed for single-mode graphs. One solution to this problem is to adapt
single-mode for a system by using domain-specific knowledge to, for example, generate
and analyse appropriate single-mode subgraphs of the system. An alternative approach is
to reduce the system to a single-mode graph by transforming one of the node types into a
relationship (e.g., transforming the researchers and papers network by removing papers
and transforming all of the edges between researchers and papers to edges between
researchers). Generally, even in situations where multiple types of nodes or edges would
provide a more accurate representation, a simple single-mode graph representation will
still provide useful insights.
 Question Local metrics Global metrics
How tightly are the entities clique, n-clique, k- clustering
(nodes) grouped together (by plex, k-core coefficient
edges)?
How many different influences degree, in-degree, degree
does an entity receive? out-degree distribution
How many other entities does it
influence?
How long does communication shortest-path diameter, radius
between nodes take? length
Do some nodes in a network play bridges, pivots degree
an important role in connecting distribution
the whole network?
Table 1: Characterising metrics by the questions that they address.

In situations where the graph must represent richer information than simple nodes and
edges will allow, it is possible to have nodes and edges with attached properties. Some
common examples of properties that may be attached to nodes are the location
information of a node in a transportation network or the activation value of a gene in a
genetic regulatory network; a link's properties may include a weight, representing the
degree of influence a gene exhibits upon the activation of another, or the strength of a
relationship in a social network. Some network types, such as transportation networks,
have sophisticated techniques for dealing with certain attached properties, such as
distance between nodes and flow through edges. In general, however, attached properties
will require domain-specific tuning of generic metrics and techniques if such properties
are to be taken into account.

Obtaining metrics through graph-theoretic analysis

As outlined above, the appropriate analysis of a graph-based representation depends on
the properties or features of the problem that are being investigated. When looking at the
natural occurrence of global stability in a genetic regulatory network, investigating node
DEGREE DISTRIBUTIONS and comparing these distributions with those of a SCALE-FREE
ARCHITECTURE may be useful; if the targets of inquiry are organisational groups in a social
network, clique analysis would be appropriate. Having identified the different structural
properties that may be of interest in a particular problem, it is subsequently necessary to
find a suitable metric with which to investigate these features. There is a list of example
questions based on structural characteristics of a graph, along with some suggestions for
which metrics to use, in Table 1.
Given that there are a set of expected local and/or global structural characteristics to be
investigated, selecting appropriate metrics involves identifying those that are likely to
provide information about the interesting structures. As most structural characteristics can
be measured in many different ways, deciding on appropriate metrics may require a
combination of domain-specific knowledge and trial and error. A more complete list of
metrics, along with the graph structures that they describe, is given in Table 2.
 Metrics Structure
degree, in-degree, out- importance of a node based on how connected it is
degree
degree distribution set of related properties, such as average shortest
path length, probability of creating disconnected
components through node/edge removal
shortest path length distance between two nodes, degree of influence
of nodes on each other
clique, n-clique, n- identification of highly interconnected subgraphs
clan, k-plex, k-core
hub identification of highly connected nodes
pivots, cut-points identification of nodes crucial to keeping the
graph connected
bridges identification of edges crucial to keeping the
graph connected
node-connectivity, how much damage the graph can take before
line-connectivity becoming disconnected
centralization how much the graph centres on a single node or
group of nodes
betweenness centrality, importance of a node based on its relationship to
closeness centrality other nodes in the network
Table 2: Some commonly used metrics and their objects of inquiry.

Case study: metrics in a genetic regulatory network

Consider a simple genetic regulatory network represented by a graph, consisting of a set
of nodes and a set of edges. The nodes in this graph are taken to represent genes, though
other regulatory elements such as RNA may also be represented as nodes. Depending on
the application, these could be represented as the same type of node for simplicity, or as a
different type of node (resulting in a multi-modal graph). Edges in the graph represent
regulatory relationships between genes; an edge from node A to B is taken to mean that
gene A regulates gene B. Whether this regulation occurs directly through proteins
produced by gene A, or whether the regulation is indirect (e.g., a protein from gene A
interferes with a protein from gene C binding to B) is not specified. Because regulatory
relationships are not bidirectional, the resulting graph is directed.
Assuming that the goal of the investigation is simply to better understand the interactions
and dependencies in the network (e.g., for purposes of drug-targeting, etc.), there are
several standard metrics that may be applied to the network. For formal definitions of
these metrics, see the online reference material.
Degree distribution (global): Does the graph have a SMALL-WORLD ARCHITECTURE or a
SCALE-FREE ARCHITECTURE? That is, are there a very few highly connected nodes which
may be central to understanding the network's behaviour? A related local metric is
the DEGREE of a node (or IN-DEGREE and OUT-DEGREE in a directed graph).
In-degree and out-degree (local, directed): Which nodes influence many other
nodes (i.e., have a high out-degree) and which are influenced by many other nodes
(i.e., have a high in-degree)? Altering a node with a high out-degree is likely to
have a catastrophic effect on the network, because the node is interacting with so
many other nodes, whereas a high in-degree may only indicate that the 'decisions'
made by the node require many different factors. A related global metric is the
DEGREE DISTRIBUTION.

Clustering coefficient (global): Broadly speaking, how closely clustered are nodes
in the network?
Cliques (local): If the network appears to be modular, do cliques (or similar
measures) identify interesting sub-components of the network? This metric can help
to identify functionally related genes/proteins in the network.
Bridges/pivots (local): If there are cliques, how do they communicate? Are there
specific nodes or edges in the network that join multiple cliques together? This
metric may identify nodes that are not highly-connected, but are nevertheless
crucial to the network's operation.

Summary: steps in representing a problem as a graph

● How will problem features correspond to graph nodes and edges?
● Will there be multiple types of nodes and/or edges?
● Will nodes and/or edges have attached properties?
● Does the problem have structural or structure dependant properties?
(e.g., systems of interacting agents)
● What kind of global structural properties might the problem have?
● What kind of local structural properties might highlight important
features of the problem?
● How can the system's structural properties be measured?
● What metrics or techniques can be used to measure global structural
properties?
● What metrics or techniques can be used to measure local structural
properties?
Tools for analysing graphs
While there are relatively few freely available programs for graph-based analysis, the
available programs are of a generally high quality, and are complete enough for a wide
range of analysis and visualisation tasks. In situations where such programs are not
sufficient – either because they cannot perform the required analysis, because they are not
fast enough or because they cannot be used without user intervention – there are freely
available libraries for several common programming languages. These libraries provide a
wide range of graph-based data structures and algorithms that can significantly reduce
hand coding. For a summary of available tools, see Table 3.

Programs

Pajek
URL: http://vlado.fmf.uni-lj.si/pub/networks/pajek/
Additional resources: http://vlado.fmf.uni-lj.si/pub/networks/pajek/apply.htm
Pajek is a program for the analysis and visualisation of large graphs. While it can be
intimidating at first glance, it is a highly capable package that is freely available, in active
development and has a large user community – particularly in the field of social network
analysis.

Tool Type License Details

Pajek Program Free General graph analysis tool
providing an excellent range of
metrics and excellent graph
visualisation capabilities, but can
be unintuitive to use.
UCINET Program Commercial Social network analysis tool
providing a variety of metrics in
a user-friendly manner.
Matlab Program / Commercial Matrix-based development tool,
Library recommended only if already
familiar.
Boost Library Free and C++ graph library providing
Graph open source graph data structures and many
Libraries (BSD-like) common graph algorithms.
Graph Library Free only for C++ graph library providing
Template academic use graph data structures and some
Libraries common graph algorithms.
JUNG Library Free and Java graph library providing
open source multiple graph data structures,
(BSD) many common graph algorithms
and visualisation tools.
Table 3: A list of available tools
Pajek is an excellent tool for visualising graphs and performing many different kinds of
standard analyses. It includes tools for analysis using clusters, cliques, components,
various partitioning schemes, general graph properties such as diameter and radius, node
degrees and many other features. The manual provides details about the available
functionality and its use.
Although Pajek is a MS Windows program, it also runs under Unix-based operating
systems via the Wine project (http://www.winehq.org/). Pajek works very well when used
in this way, with few noticeable problems.

UCINET
URL: http://www.analytictech.com/ucinet.htm
UCINET is a package developed for social network analysis and contains a range of
network analysis tools. There is a freely available demonstration version, with registration
required after 30 days. The interface for UCINET is much more polished than that of
Pajek, but makes heavy use of terms from social network analysis.
It is a good tool for graph analysis, and a reasonable tool for graph visualisation. The
range of analysis tools included is not as wide as that of Pajek, but the available tools are
easier to use. It is accompanied by a high quality manual and a PDF version of a good
(free) social network analysis textbook.
Although UCINET is a MS Windows program, it also runs under Unix-based operating
systems via the Wine project (http://www.winehq.org/). UCINET works reasonably well
when run in this way – the analysis tools generally work well, but there are problems with
graph drawing.

Matlab
URL: http://www.mathworks.com/
User community: http://www.mathworks.com/matlabcentral/
While Matlab is not designed for working with graph-based representations, it is capable
of efficiently and simply dealing with connectivity matrix representations of graphs.
Although not a recommended starting point, researchers familiar with Matlab may find its
capabilities sufficient for a variety of graph analyses. There are some user-contributed
libraries available to support the use of Matlab for graph analysis that can be found
through the user community. However, there are no official packages available, and the
tools available through the user community are not as well developed as similar packages
for C++ or Java (see below). Some of the available user-contributed packages are:
● GrTheory – a graph theory toolbox.
http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objectId=4266
&objectType=file
This toolbox provides several graph algorithms such as minimum spanning tree,
minimum vertex cover and all-pairs shortest path, but has no internal graph
representation; instead, it uses simple node and edge vectors.
● Toolbox Graph – a toolbox for graph manipulation and plotting
http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objectId=5355
&objectType=file
 This toolbox includes few algorithms, but has an internal graph representation with
support for building several graph types, as well as having some support for graph
visualisation.

Libraries

Boost Graph Libraries (C++)

URL: http://www.boost.org/
The Boost Graph Libraries are a generally high-quality implementation of basic graph
data structures and algorithms, and are an excellent starting point for hand-coded
simulations. The libraries make extensive use of C++ template functionality, which
makes extending the libraries possible. However, the complexity of the template-based
code makes such extension difficult for programmers without significant experience in
C++ templates.

Graph Template Library (C++)

URL: http://www.infosun.fmi.uni-passau.de/GTL/index.html
The Graph Template Library provides basic graph data structures that integrate well with
the C++ Standard Template Libraries. However, its functionality is limited relative to the
Boost Graph Libraries. In addition, the license under which the library is covered
restricts commercial use of the library.

Jung (Java)
URL: http://jung.sourceforge.net/
Jung is an actively developed Java graph library with a promising set of graph algorithms
available, and a competent visualisation framework. It is easily extensible through
standard Java mechanisms. It also imports and exports Pajek data files.

Bibliography
Diestel, Reinhard (2000). “Graph theory” Springer: NY.
Hannenman, Robert (2001). “Introduction to Social Network Methods” Available from
http://faculty.ucr.edu/~hanneman/networks/nettext.pdf and included in the UCINET
package.
Strogatz, Stephen H. (2001). “Exploring Complex Networks” Nature 420, pp. 268-276.
Wasserman, Stanley and Faust, Katherine (1994). “Social Network Analysis : Methods
and Applications” Cambridge University Press: Cambridge, NY.

Glossary

BETWEENNESS A measure of centrality associated with a NODE. The total number

CENTRALITY of shortest PATHS between every possible pair of nodes in a GRAPH
that pass through the given node. See also CLOSENESS CENTRALITY.
BRIDGE An EDGE that will, if removed, cause a GRAPH (or subgraph) that
was connected to become disconnected. See also PIVOT.
CENTRALIZATION A measure of how much a GRAPH centres on a particular set of
NODES. Generally, the ratio between the actual and the maximum
possible sum of differences between the centrality of the most
central node and that of all other nodes in the graph. This metric
will be different depending on which measure of node centrality
(e.g., CLOSENESS CENTRALITY, BETWEENNESS CENTRALITY) is used.
CLIQUE A set of NODES in a GRAPH such that each node in the set is
connected to every other node in the set. See also N-CLIQUE, N-
CLAN, K-PLEX, K-CORE.

CLOSENESS CENTRALITY An inverse measure of centrality associated with a NODE. The sum
of the SHORTEST PATH LENGTHS between a given NODE and all other
nodes in the GRAPH. See also BETWEENNESS CENTRALITY.
CLUSTERING A measure of the closeness of NODES in a GRAPH. The average of
COEFFICIENT the clustering coefficient of all the nodes in a graph, where the
clustering coefficient of a node is the probability that two
neighbours of the node are connected to each other.
CUT-POINT See PIVOT.
DEGREE The number of other nodes that a given node is connected to by
EDGES. For directed graphs, see IN-DEGREE and OUT-DEGREE.

DEGREE DISTRIBUTION The distribution of NODE DEGREES in a GRAPH, visually represented

by a frequency histogram of degrees. For example, see SCALE-FREE
ARCHITECTURE.

DIAMETER The longest PATH between any two NODES in a GRAPH.

DIRECTED GRAPH A GRAPH containing directed EDGES is a directed graph. In a
directed graph, the relationships between NODES are directed –
they have a start node and an end node. See UNDIRECTED GRAPH.
EDGE An edge is a fundamental component of a GRAPH, generally used
to represent a relationship in a problem. An edge connects two
NODES together. An edge may be directed, in which case it has a
specific source node and a target node, or undirected.
EDGE-CONNECTIVITY The minimum number of EDGES needing to be removed in order to
disconnect a given GRAPH. See also NODE-CONNECTIVITY.
GRAPH A graph consists of a set of NODES and a set of EDGES that connect
nodes together. Graphs are useful problem representations due to
their capability to highlight structural characteristics of the
interactions between elements of the graph.
HUB A NODE in a GRAPH with a high DEGREE. Sometimes taken to be a
node that is connected to every other node in the graph.
IN-DEGREE The number of other NODES that a node receives input from. Also,
the number of EDGES of which this node is the target. For
undirected graphs, see DEGREE.
K-CORE A set of NODES in a GRAPH such that each node in the set is
connected to at least k other nodes in the set. (i.e., in a 2-core,
every node is connected to at least 2 nodes in the set) See also
CLIQUE, K-PLEX.

K-PLEX A set of NODES in a GRAPH such that each node in the set is
connected to all but k other nodes in the set. (i.e., in a 2-plex,
every node is connected to at least all but 2 nodes in the set) See
also CLIQUE, K-CORE.
N-CLAN A set of NODES in a GRAPH such that each node in the set is
connected to every other node in the set by n or fewer EDGES, all of
which are between two nodes in the set. (i.e., in a 2-clan, every
node is connected by no more than 2 edges in the set) See also
CLIQUE, N-CLIQUE.

N-CLIQUE A set of NODES in a GRAPH such that each node in the set is
connected to every other node in the set by n or fewer EDGES. (i.e.,
in a 2-clique, every node is connected by no more than 2 edges)
See also CLIQUE, N-CLAN.
NODE A node is a fundamental component of a GRAPH, generally used to
represent an entity in a problem. Nodes are connected together by
EDGES.

NODE-CONNECTIVITY The minimum number of NODES needing to be removed in order to

disconnect a given GRAPH. See also EDGE-CONNECTIVITY.
OUT-DEGREE The number of other NODES that a node sends output to. Also, the
number of EDGES of which this node is the source. For undirected
graphs, see DEGREE.
PATH A series of unique NODES describing a particular traversal of a
GRAPH.

PIVOT A NODE that will, if removed, cause a GRAPH (or subgraph) that was
connected to become disconnected. See also BRIDGE.
RADIUS The shortest longest PATH between the most central NODE in the
GRAPH, and all the other nodes in the graph

SCALE-FREE A GRAPH with a scale-free architecture is one that has a DEGREE

ARCHITECTURE DISTRIBUTION in which the probability P(k) of a NODE having a
-g
DEGREE of k is described by an inverse power law, P(k) = k .
Graphs with scale-free architectures have been shown to be robust
to random damage, and have a very small RADIUS.
SHORTEST PATH The number of EDGES in the shortest path that connects the two
LENGTH given NODES in the GRAPH.
UNDIRECTED GRAPH A GRAPH containing undirected EDGES is an undirected graph. In an
undirected graph, the relationships between NODES are reflexive –
they are between nodes, rather than being from one to another.
See DIRECTED GRAPH.