Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Exploring Statistical and Population Aspects of Network Complexity

  • Frank Emmert-Streib ,

    v@bio-complexity.com

    Affiliations Computational Biology and Machine Learning Lab, Center for Cancer Research and Cell Biology, School of Medicine, Dentistry and Biomedical Sciences, Queen’s University Belfast, Belfast, United Kingdom, Department of Biostatistics and Computational Biology, Dana-Farber Cancer Institute, Harvard School of Public Health, Boston, Masachusetts, United States of America

  • Matthias Dehmer

    Affiliation Institute for Bioinformatics and Translational Research, UMIT, Hall in Tyrol, Austria

Exploring Statistical and Population Aspects of Network Complexity

  • Frank Emmert-Streib, 
  • Matthias Dehmer
PLOS
x

Abstract

The characterization and the definition of the complexity of objects is an important but very difficult problem that attracted much interest in many different fields. In this paper we introduce a new measure, called network diversity score (NDS), which allows us to quantify structural properties of networks. We demonstrate numerically that our diversity score is capable of distinguishing ordered, random and complex networks from each other and, hence, allowing us to categorize networks with respect to their structural complexity. We study 16 additional network complexity measures and find that none of these measures has similar good categorization capabilities. In contrast to many other measures suggested so far aiming for a characterization of the structural complexity of networks, our score is different for a variety of reasons. First, our score is multiplicatively composed of four individual scores, each assessing different structural properties of a network. That means our composite score reflects the structural diversity of a network. Second, our score is defined for a population of networks instead of individual networks. We will show that this removes an unwanted ambiguity, inherently present in measures that are based on single networks. In order to apply our measure practically, we provide a statistical estimator for the diversity score, which is based on a finite number of samples.

Introduction

Complexity is a general notion that triggered a large number of studies in a variety of different fields, ranging from biology, chemistry and mathematics to physics [1][9]. Despite this attraction, up-to-now a generally accepted description of the complexity of an object that would allow the establishment of a quantitative measure for its characterization is still absent. Probably the best studied objects with respect to the characterization of their complexity are one- and two-dimensional strings or symbol sequences. For such objects, many approaches have been suggested to define or assess complexity quantitatively [3], [8], [10][18]. However, an intrinsic problem of any complexity measure is that there are alternative ways to perceive and, hence, describe complexity leading inevitably to a multitude of different complexity measures [19]. For example, Kolmogorov complexity [2], [3], [8], [20] is based on algorithmic information theory considering objects as individual symbol strings, whereas the measures effective measure complexity (EMC) [16], excess entropy [21], predictive information [22], thermodynamic depth [17] or statistic complexity [14] relate objects to random variables and, hence, are ensemble or population based.

In the context of networks, graph complexity measures have been suggested to investigate the complexity of chemical graphs representing molecules and chemical compounds [23][25]. Different types of graph complexity measures have been developed which can be broadly divided into information-theoretic and non-information-theoretic measures. Because so far it is largely unclear what structural features of a network to emphasize, hierarchical approaches for the chemical complexity consisting of several hierarchical levels of molecular complexity have been developed. One of the first attempts was due to Bertz [26] developing a hierarchical model containing both topological (i.e., branching, rings, multiple bonds) and non-topological (molecular size, symmetry, functionality, elemental composition) features; for a detailed discussion see [25]. Later, Bonchev and Polansky [27] furthered this system and described the total complexity of a chemical system by a vector approach. The components of this vector represent various features of complexity, e.g., the system size, graph topology, physical nature, metric of a system and its symmetry [27].

Also for general networks there are many network complexity measures that have been suggested [24], [28]. Many of these are based on information-theoretic principles [29][31]. A classical, non-information-theoretic approach is the so-called combinatorial complexity, introduced by Minoli [32]. This measure represents a monotonically increasing function of the factors which contribute to the complexity of a network, e.g., the number of vertices and edges, vertex degrees, multiple edges, cycles, loops, and labels [33]. Other techniques rely on determining particular substructures in graphs [24], [28]. Also Constantine et al. [34] defined the complexity of a graph to be the number of its containing spanning trees. An operator approach has been developed by Jukna [35] who defined graph complexity as the minimum number of union and intersection operations required to obtain the entire set of its edges starting from star graphs. Approaches to define the complexity of graphs based on Kolmogorov’s complexity paradigm [3] can be found in [36], [37]. Particularly, Bonchev [37] compared the Kolmogorov complexity of a graph with other measures and tackled the problem whether all these techniques can detect branching in graphs.

The major purpose of this paper is to introduce a network measure, called the network diversity score and to demonstrate that this measure allows to categorize networks with respect to their structural complexity. Specifically, we demonstrate that the diversity score allows to distinguish ordered, random and complex networks from each other. Further, we study 16 additional network complexity measures and find that none of these measures has similar good categorization capabilities with respect to the structural complexity of networks. In contrast to many other measures suggested so far, the network diversity score is different for a variety of reasons. First, our score is multiplicatively composed of four individual scores, each assessing different structural properties of a network. That means our overall score reflects the structural diversity of a network. Abstractly, this may be seen as the dimension of the complexity of a network. Second, our score is defined for a population of networks instead of individual networks. We will show that this removes an unwanted ambiguity, inherently present in measures that are based on single networks. To enable a practical application of the network diversity score we provide a statistical estimator for this score that is based on a finite number of networks sampled from the underlying population of networks.

This paper is organized as follows. As the definition for a structural complexity of networks suffers from similar problems as for one-dimensional symbol strings, several heuristic criteria have been proposed, with which a complexity measure should be conform [25], [27]. In order to clarify what we mean by a complex network we provide in section ‘Characterizing the complexity of networks’ a description of this, on which we rely in this paper. Then we describe 16 network complexity measures used for our analysis and characterize their computational complexity. In order to present the network complexity measures used in this paper, we roughly categorize them into two classes: information-theoretic and non-information-theoretic measures. Clearly, each group can be further subcategorized. For instance, we could subsume the class of pure distance-based and eigenvalue-based measures under the category of non-information-theoretic measures. As known, information-theoretic graph complexity measures [23], [38] rely on inferring a probability distribution by taking structural features of a graph into account. More precisecly, so-called partition-based and non-partition-based measures can be derived by using Shannon’s entropy, see [23], [39]. Other graph entropy measures based on using subgraph-relations can be found in [28]. Non-information-theoretic complexity measures are mostly based on transforming simple graph invariants such as vertex degrees and distance-based quantities [40] into real numbers [41], [42]. For instance, the first zagreb index [41], [42] transforms vertex degrees into a positive measure for characterizing the structure of the graph. Another class of non-information-theoretic complexity measures is based on deriving subgraphs and then transforming them into measures finally leading to a graph complexity measure, see [28]. In section ‘Network diversity score’ we define our measure and clarify conceptual differences to other approaches. In the results section we investigate all 17 network measures for a variety of different settings and compare them with each other. The paper finishes with a ‘Conclusion’ section, summarizing the obtained results.

Methods

In this section we, first, provide a characterization for the complexity of networks as used in this paper. Then, we describe 16 network complexity measures we are using in our analysis and characterize their computational complexity. Thereafter, we introduce a new complexity measure, called network diversity score (NDS), and provide a motivation for its definition.

Characterizing the Complexity of Networks

As outlined in the introduction, so far there is no universally accepted definition of complexity available that would be applicable to general objects, including networks. However, it is generally believed that a complexity measure should be capable of distinguishing complex objects from random and ordered objects. For objects generated by a physical process this complexity characterization has been given in [4], [19]. However, also for the complexity of biological systems similar assertions have been made [43]. In the following we adopt this perspective. Figure 1 A provides a visualization of this characterization, placed in the context of networks. In this figure the x-axis corresponds to an one-dimensional variable that represents networks G from the network space , and the y-axis gives the value of the complexity measure . Here, the variable q is assumed to represent networks of a similar type smoothly. That’s why certain regions of the x-axis have been labeled as, ordered, complex or random. Concrete examples for such a variable is Langton’s [44] for one-dimensional cellular automata or the mean connectivity K in random boolean networks [45].

thumbnail
Figure 1. A: Visualization of the properties of a complexity measure with respect to different networks.

B: Alternative complexity measure with different characteristics.

https://doi.org/10.1371/journal.pone.0034523.g001

It is important to clarify the relation between three different entities: the network G, the variable q representing a network and the complexity measure M. A network is an abstract object which possesses a multitude of different properties, e.g., number of nodes, degree distribution, mean path length between all nodes, to mention just a few. For this reason, a network is not easily quantifiable by a singe variable because a mapping, , is usually not unique. For example, if we identify with the (global) clustering coefficient of network G [46], then there are many networks that have the same value of q. For this reason, when one maps a network G to q, the value of q represents actually a set of networks that map to the same value of q, i.e., with . Similar arguments hold when we map a network to its complexity value, i.e., . Also in this case, usually, many networks map to the same complexity value, with . It is interesting to note that after networks have been identified as complex, random or ordered, by using the complexity measure M, the entity q can serve itself as a complexity measure, if it exhibits a smoothness property with respect to the underlying networks. Here, smoothness means that similar networks lead to similar values of q. This smoothness property allows the identification of continuous regions (intervals) of q values, which represent specific types of networks, as shown in Fig. 1 A.

The particular problem we want to study in this paper differs from the above. Instead of using a complexity measure M to categorize networks into the groups complex, random or ordered, we assume that such a categorizations for the networks is already known. From the above discussion we know that if we find a smooth measure , representing sets of networks that assigns to similar network types, similar values of q, then q can serve as a complexity measure. That means for networks that are labeled according to certain categories they belong to and a measure q, one can quantitatively assess the quality of such a measure with respect to the given labels of the networks. Hence, by using the knowledge of the labeling of different networks we can investigate the categorization abilities of a measure q.

In Fig. 1 B we show an alternative behavior of a complexity measure in dependence on networks. In this case, we called the values on the y-axis ‘score’ and not complexity measure because here a score for complex networks does not lead to the highest possible values but to intermediate values. However, the advantage of such a score, compared to the ones illustrated in Fig. 1 A, is that it allows to discriminate between all three network types, complex, ordered and random networks, considering the score of the networks only. Hence, there are three continuous regions of values of the score that allow to distinguish the three types of networks unambiguously. Other configurations may be possible and helpful, however, in the following, we base our analysis on this basic characterization of complexity and apply it to networks. As our numerical results will demonstrate, the principle behavior of the score sketched in Fig. 1 B is of practical relevance for our analysis (see Fig. 8 and its discussion).

Definition of Complexity Measures

In the following we provide a brief description of the complexity measures we are using in our study. We denote by G a network having vertex set V and edge set E. The number of vertices is and the number of edges . Table 1 gives an overview of the 16 complexity measures we use.

thumbnail
Table 1. Overview of the network complexity measures we use in our analysis.

https://doi.org/10.1371/journal.pone.0034523.t001

Information-theoretic Complexity Measures.

A variety of entropic measures determining their structural information content have been developed to characterize networks structurally [38]. The following measures are based on Shannon’s entropy.

  • Topological information content:

One of the first measures was the topological information content introduced by Rashevsky [58] given by(1)

Here, denotes the number of topologically equivalent vertices in the i-th vertex orbit of G and k is the number of different orbits. is a measure of symmetry in graphs. This measure vanishes for a fully symmetric graph such as regular graphs and attains its maximum value for asymmetric graphs. Importantly, Trucco [59] also investigated this measure and Mowshowitz [56] generalized it to determine the structural information content of graphs and studied mathematical properties thereof [56], [60], [61].

  • Bertz index:

A more general graph complexity measure is due to Bertz and expresses the total structural information content of a graph:(2)

X is an arbitrary graph invariant such as its vertices, edges, degrees etc. refers to its cardinality. For example, if X corresponds to the vertices of a network than corresponds to the number of vertices. If we choose , we get(3)

as special case.

  • Bonchev-Trinajstić index:

By defining weighted probability schemes, one generalizes classical measures of Rashevsky and Mowshowitz [56], [58], see Eq. 1. A special measure thereof is given by(4)

This measure is based on the Wiener-Index [57],(5)

Note that the Wiener index is the sum of all distances in a graph G. The distances can be computed by using Dijkstra’s algorithm or any other method for calculating shortest paths in a graph [62], [63]. Here, is the diameter of network G and is the number of the shortest paths having length i.

  • Information-theoretic complexity measure based on information functionals:

The following measure belongs to a family of graph entropy measures based on using information functionals [39]. A special measure thereof is the degree-degree association index as it is based on the special information functional , see [52]. The functional is defined by(6)

The detailed explanation and definition can be found in [52]. The degree-degree association index is defined by(7)

is a scaling constant. Note that is not based on determining partitions of graph elements in a classical sense (such as ) as probability values are assigned to each vertex of G.

  • Offdiagonal complexity:

To define Offdiagonal complexity () [54], let be the vertex-vertex link correlation matrix, see [54]. denotes the number of all neighbors possessing degree of all vertices with degree i [28]. stands for the maximum degree of G. If one defines [28](8)

and(9)

can be defined by [28](10)

  • Spanning tree Sensitivity:

The following measure is based on determining substructures in graphs. The spanning tree sensitivity [28] is defined by(11)with , , and being an ordered list of all k different . is the number of spanning trees in the graph minus the number of spanning trees of the subgraph with the edge deleted. Analogously, the spanning tree sensitivity differences measure is defined as(12)with , where is the ordered list of all unique differences .

Non-information-theoretic Complexity Measures.

Non-information-theoretic complexity measures for networks can be defined by using arbitrary graph invariants such as distances between nodes or their degrees. In the following, we describe some important measures which have already been used in a variety of different disciplines.

  • Balaban J :

The Balaban J index is defined as [42], [47](13)

denotes the sum of distances from vertex to all other vertices, i.e.,(14)whereas D is the distance matrix containing the shortest path lengths between all vertices measured by the Dijkstra distance [63] and is the cyclomatic number [64].

  • Complexity index B :

The complexity index B is a more recently developed measure due to Bonchev [24]:(15)

where(16)

Here, is the degree of a vertex .

  • Efficiency:

Latora et al. [49], [50] developed a measure called the Efficiency complexity of a graph G. Starting from(17)

expressing the arithmetic mean of all inverse path lengths and(18)

the Efficiency complexity yields to(19)

  • Mean distance deviation:

In general, distance-based measures are straightforward to calculate with polynomial time complexity [62]. Hence, a variety of distance-based indices have been developed to characterize networks based on their topology [40], [65]. The mean distance deviation introduced by Skorobogatov and Dobrynin is defined as [40], [42]:(20)

where(21)

and(22)

  • Normalized edge complexity:

The normalized edge complexity using the elements of the adjacency matrix has been introduced by Bonchev [24]:(23)

where(24)

Here, denotes the entry in the i-th row and j-th column of the corresponding adjacency matrix A.

  • Randić connectivity index:

The Randić connectivity index [55](25)has been sucessfully used as branching index. Also, R has been explored extensively, e.g., bounds and other extremal properties have been invesitagted in an interdisciplinary manner [66].

  • Wiener index:

One of the first structural graph decsriptors was the Wiener-Index [57],(26)

denotes the shortest distance between and .

  • Zagreb index:

A classical degree-based index based on the vertex degree is the first Zagreb index [41], [42] defined as(27)

is just the sum of the vertex degrees of G.

Eigenvalue-based Measures.

By determining the eigenvalues of graph-theoretical matrices such as the adjacency matrix or the Laplacian, various measures can be obtained [51], [67].

  • Graph energy:

Gutman [51] defined the sum of the absolute values of eigenvalues of the adjacency matrix of a graph and called the resulting quantity graph energy.(28)where are the non-zero eigenvalues of the adjacency matrix of G.

  • Laplacian energy:

Instead of using the eigenvalue of the adjacency matrix of a graph, several other graph-theoretical matrices can be used. By using the Laplace matrix, we obtain the laplacian energy [42] defined by(29)

Here are the eigenvalues of the adjacency matrix and those of the Laplacian matrix of the graph.

Computational Complexity

Calculating the complexity of networks can be computationally intense and many algorithms are even NP-complete [68]. For instance, determining the automorphism group of a general graph to compute the graph entropy measure is computationally demanding as the computational complexity can be exponential [69]. In contrast, the time complexity of some information-theoretic graph complexity measures such as B, OdC, and is polynomial, see [70]. Particularly the time complexity of the Bonchev-Trinajstić index and the degree-degree association index is as we need to calculate all shortest paths between all vertices in the graph leading to . Similar statements [28], [70] for the time complexity of J, and can be obtained as the complete distance matrix needs to be calculated. Simple topological network measures, such as the Wiener and Randić index also possess polynomial time complexity as their calculation rely on matrix computations based on graph invariants.

The time complexity of determining the zeros (eigenvalues) [71] of graph polynomials [51] such as the characteristic or distance polynomial is polynomial too. For instance, by using the adjacency matrix to calculate the characteristic polynomial of a graph, we obtain its eigenvalues in polynomial time. From this, measures such as the graph energy E and the laplacian energy LE can be calculated efficiently.

Network Diversity Score

In the following we define a network measure we call the network diversity score (NDS). Our score is based on 4 variables:(30)(31)(32)(33)

Here, M is the number of modules in a network G and n is the number of vertices of this network. The vector contains the size of the modules, i.e., gives the size of the i-th module, which corresponds to the number of nodes in this module. To identify the modules in a network we use a method called Walktrap [72] which finds modules based on random walks similar to [73], [74]. An advantage of this method over many others is its efficient computational complexity, given by (e: number of edges, n: number of vertices). The vector in Eqn. 32 represents the eigenvalues of the Laplace matrix L of network G [75], whose components are defined by(34)

Here, is the degree of node i in G. Finally, and correspond to the number of motifs of size 3 and 4 found in network G [76]. That means is the number of different motifs one can find in G having i nodes.

Based on the above four variables, we define the individual diversity score for a network G by(35)

We call this measure individual diversity score because it can be calculated for a single network G. The individual diversity score assesses one network G and assumes values in . Based on we define the network diversity score (NDS), , for a population of networks by(36)

Here, denotes the population of networks that belong to the same network model and is a probability density over this population. For example, this could correspond to the random network model generated with the Erdös-Réyni model [77], [78]. Or it could be the set of all scale-free networks generated with the preferential attachment algorithm [79], [80]. Or the population could contain all networks that have the same degree, e.g., a lattice with periodic boundary conditions. That means the population of networks can be either defined by a stochastic process that generates the networks in the population or by structural properties of the networks themselves.

In order to obtain an approximation of the measure , which can be applied to a finite set of networks, we define the network diversity score for a sample of size from the population by the estimator,(37)

Assuming that the S networks are independently sampled from the population than, according to the central limit theorem [81],(38)

For our numerical investigations we use the estimator given in Eqn. 37.

The diversity score represents the idea that a network is a high-dimensional object. Specifically, we consider the 4 variables and as important. The variable provides information about the module density of a network. For complex networks we would expect to find more modules than for random networks because modules are an expression of a general organizational principle of a network. The variable is a rate about the growth of motifs within a network. From numerical results we observed that ordered networks have the highest, complex network have intermediate and random networks have the lowest values of . The variable is similar to a CV (coefficient of variation) value which measures the variability of network sizes with respect to the mean size of a module. Random networks are expected to have a low variability of module sizes but also a low mean module size whereas complex networks should have a higher variability of module sizes but also a higher mean module size. The variable is similar to but for the eigenvalues of the Laplace matrix L. We studied many combinations of these 4 and other variables and found from numerical investigations that the individual density score in Eqn. 35 results in the best separation of random, complex and ordered networks.

Motivation for the network diversity score.

The underlying rational of our measure is based on the following observations. First, studies investigating the complexity of various types of objects, e.g., one-dimensional strings, led to the introduction of a large number of different complexity measures. However, up-to-now there is no general agreement that the right measure is among the introduced ones. For networks, we are facing a similar situation that may be potentially even more severe. For this reason, we are proposing a composite measure that is not just based on the evaluation of one structural principle, but on the combination of several ones. Hence, their combinatorial usage abates the need for each individual measure to represent the right complexity measure. In the results section, we will numerically demonstrate that such a composite measure leads in fact to very good results.

A second reason that motivated us to introduce our measure is best described by the following illustration. Suppose, one defines networks as ‘random’ when they have been generated with the random network model, suggested by Erdös-Réyni and Gilbert [77], [78], and as ‘complex’ when they have been generated with the preferential attachment algorithm [79], [80]. Then, there exists a non-vanishing probability to generate a random network with the random network model that is also complex. However, this is counter intuitive. Let us consider an example for this problem. Suppose, a network has been generated with the random network model and a second network has been generated with the preferential attachment algorithm. Then, with a certain probability, (with the meaning ) holds, because the random network model can, in principle, generate all possible network structures. More precisely, if the undirected network contains e edges (denoted by ) and n vertices then it contains missing edges (non-edges). That means the probability, w, for the random network model to generate a particular network with e edges is given by(39)

Here, is the probability to have e edges in and is the probability to have non-edges in . That means, assigning a complexity value to individual networks leads to a loss of the unique connection between the complexity of the network and the underlying network model that generated this network. This is visualized in Fig. 2 A. In this figure, w corresponds to the probability that the random network model generates a complex network . Starting from the complexity value of a network, right hand side of the figure, one sees that it is possible to conclude that has been either generated with a random network model or with a complex network model. For reasons of simplicity, we used in the above explanation only two network models, however, an extension to more models is straight forward, but makes the explanations more laborious. It should be clear that in such an extended scenario, the potential for an ambiguity between the complexity of individual networks and the network generating models is even amplified.

thumbnail
Figure 2. Connection between network model, networks and a complexity measure assessing either the complexity of individual networks (A), a population of networks (B) or a sample of networks (C).

https://doi.org/10.1371/journal.pone.0034523.g002

In order to avoid this problem, we base our network score on the principle visualized in Fig. 2 B. Due to the fact that the complexity is assessed for a network population, generated by a network model, there is no confusion with respect to the underlying network model that generated the population, because the complexity measure can rely on the information provided by the whole population and not only by an instance thereof. Practically, we approximate such a population measure by using a finite sample of networks, as shown in Fig. 2 C. For a finite sample consisting of S networks, there is also a non-vanishing probability to result in an ambiguous connection between the complexity and the underlying network model that generated the network sample, visualized in Fig. 2 C. However, this probability is only , compared to w for a complexity measure relying on a single network. In the limit for this probability goes to zero and model C becomes model B for any . Hence, using a sample of size S reduces the potential for an ambiguity leading to a miscategorization by a factor of . For example, if and the sample size is only than this factor is already .

We would like to emphasize that the above explanations are intended as a motivation of our approach and not as a numerical analysis of the most general situation conceivable. In this respect, the probability w given in Eqn. 39 needs to be adapted for more general situations. However, regardless of its precise value, w will be always larger than zero and the principle discussion above translates seamlessly to more involved conditions. In the next section, we provide a numerical analysis for a large variety of different networks.

Results

We begin our analysis by investigating the statistical variability of the 16 network complexity measures listed in Tab. 1. In Fig. 3 we show results for 100 networks generated with the random network model [77], [78] for the parameters and . Here n corresponds to the number of nodes in a network and the parameter is the probability with which two nodes are connected by an edge. Each histogram shows the result for one complexity measure, as indicated by the name in the legend. The x-axis corresponds to the value of the respective complexity measure and the y-axis gives the frequency of observed values. It is important to note that despite the fact that all random networks have been generated for the same network parameters, n and , the resulting complexity measures do not provide identical results but fluctuate. We repeated this analysis for different parameters of the random network model and also for different network types, i.e., for complex networks. For all studied cases, we found qualitatively similar results. This reveals a common conceptual drawback of all these network measures because none of the measures is considered as a random variable. However, due to the fact that a network is sampled from an underlying population, this network varies structurally, and, hence, also the network measure, as seen in Fig. 3. That means ignoring this fact is counter productive and results in a loss of interpretability of these network measures, as will be demonstrated later in this section (see Fig. 7). As explained in section ‘Network diversity score’, a random network model is in principle capable of generating all possible types of networks, including ordered and complex networks, however, only with a certain probability. Due to the fact that all measures assess only one network, which has been randomly sampled from the underlying population of a network model, the sampled network conveys the variability of network structures of the population to the network measure itself.

thumbnail
Figure 3. Evaluation of 100 random networks generated with the random network model, and .

Each histogram shows the results for one network measure; see the legend for the name of the measure.

https://doi.org/10.1371/journal.pone.0034523.g003

In the Figs. 4 to 5 we show results for two different network models and the influence of model parameters on the 16 complexity measures. In Fig. 4 we show results for a random network model with a connection probability between nodes of (x-axis). Fig. 5 shows results for a small-world network model [82] for a rewiring probability of (x-axis). In these figures, the mean value and the standard deviation of a complexity measure (y-axis) is shown in dependence of the model parameter (x-axis).

thumbnail
Figure 4. Random network model: Dependence of the complexity measures (y-axis) on (x-axis).

https://doi.org/10.1371/journal.pone.0034523.g004

thumbnail
Figure 5. Small-world network model: Dependence of the complexity measures (y-axis) on (x-axis).

https://doi.org/10.1371/journal.pone.0034523.g005

Fig. 4 demonstrates that among the 16 complexity measures, one can observe four qualitatively different types of behavior. The four observed behavior are: (1) a monotonous increase in the complexity value (complexityIndexB, efficiency, energy, lapEnergy, randic, sTreeSens, tInfoContent, zagreb1), (2) a monotonous decrease in the complexity value (infoTheoGCM), (3) increasing complexity values followed by decreasing values (bonchev2, mDistDev, wiener), (4) decreasing complexity values followed by increasing values (balabanJ, nEdgeComplexity, offdiagonal). This indicates that different network measures have entirely different characteristics due to different structural features of the network they capture. Further, we observe that all measures, except infoTheoGCM, result in non-overlapping values for different model parameters which means that different values of lead to significantly different values of the corresponding complexity values. This is important to note since all networks generated with the random network model for different values of are random networks.

The results for the small-world network model, shown in Fig. 5, are principally different to the results shown in Fig. 4, because for different values of we obtain different network types. Specifically, we obtain ordered (), complex () and random networks (). This is different to the results for the random network model because different model parameters result always in a random network, whereas for a small-world network model, different model parameters lead to a different type of a network. Among the 16 network measures, 5 demonstrate a discriminative behavior with respect to the three different network types (balabanJ, complexityIndexB, energy, mDistDev and sTreeSens). That means these 5 measures exhibit for complex networks () noticeably different values than for ordered and random networks.

In Fig. 6 we show results about the influence of the network size n, ranging from 100 to 500 nodes, on the complexity measures. Because the type of a network does not change for a different size of the network, one would ideally expect constant values of the network measures for all different network sizes. The only measures that are approximately constant are offdiagonal and sTreeSens because their mean complexity values do not change much if taking the standard deviation of the measure into consideration. All other measures are significantly effected by the size of the networks. This hints that the size of a network is an important parameter. To simplify the following analysis, we study only networks of a fixed size.

thumbnail
Figure 6. Dependence of the complexity measures on the size n (x-axis) of small-world networks.

https://doi.org/10.1371/journal.pone.0034523.g006

So far, we studied only individual network models for a variety of different parameters these models depend on. Now, we investigate a mixture of different network models. More specifically, we generate a set, , consisting of 1500 networks, each with vertices. This set is composed of 200 ordered networks, 600 random networks and 700 complex networks. The set of complex networks is itself a mixture of scale-free networks, with different parameters of the power of the preferential attachment model , and small-world networks, with a rewiring probability of . For the set of random networks we used different parameters to connect vertices with an edge, namely, . Also, we generated random networks with the small-world model by setting the rewiring probability to 1.0. That means the resulting set of networks is heterogeneous with respect to the generation of the used networks. The median number of edges of these sets of the ordered, random and complex networks is 200 for each network type and their standard deviation is 109,43 and 60. The same data set will later be used to study the network diversity score (see Fig. 8).

thumbnail
Figure 7. Density of the 16 complexity measures for ordered (red), complex (purple) and random (green) networks.

https://doi.org/10.1371/journal.pone.0034523.g007

Application of the 16 complexity measures to leads to the results shown in Fig. 7. These figures show the probability density of the complexity values (y-axis) in dependence on the complexity values of the networks (x-axis). The three different colors correspond to ordered (red), complex (purple) and random (green) networks. The ideal behavior of a complexity measure we would like to observe is a separation of the three different network types, which means the density of the complexity values for ordered, complex and random networks should only marginally be overlapping to enable a meaningful categorization of the three network types. Considering the obtained numerical results in Fig. 7 from this perspective we find that only the offdiagonal complexity allows, at least to a certain degree, to separate the three network types from each other. The densities of all other measures do not separate at all. The problem with the density for the offdiagonal complexity is not only that it is bimodal for complex networks but also that there is still a considerable overlapping of complex (purple) and random networks (red).

thumbnail
Figure 8. Density of the diversity score for ordered (red), complex (purple) and random (green) networks.

The first row shows results for networks with nodes and the second row for nodes. The four columns correspond to the four sample sizes .

https://doi.org/10.1371/journal.pone.0034523.g008

Next, we investigate the behavior of the network diversity score, , given in Eqn. 37. In the top row in Fig. 8 we show the results for the application of the diversity score to . Due to the fact that our complexity score depends on the sample size S, the four columns in Fig. 8 correspond to four different sample sizes (). Hence, the number of different networks used for these four cases are which equals to networks. We would like to emphasize that for , the estimator gives the worst possible approximation for the density score . This case is not included to suggest it is a potential choice of S, instead, it is included to demonstrate the strength of a population effect for values of . For this reason, we highlight the difference of the case from the others, by framing the first column in Fig. 8 by a blue rectangle to indicate that it is not meant as a suggested value for the sample size.

From Fig. 8 one can see that for increasing values of the sample size S, the three network types - ordered networks (red), complex networks (purple) and random networks (green), respectively their densities become more and more separated from each other, as desired. But even for the sample size , the results for the diversity score are improved compared to the offdiagonal complexity, which was the best performing measure of all 16 network measures. The second row in Fig. 8 shows a similar analysis, however, for networks having nodes for which we generated another set of networks containing networks. For we observe an even clearer distinction of the three network types, which separate for perfectly from each other. We would like to emphasize that due to the nature of the network diversity score, which is population based, a comparison with any of the 16 network measures is uneven because none of these measures can be influenced by the sample size S. On the other hand, a sample of networks of size S contains valuable information that can be exploited to increase the discriminative abilities of a measure, as demonstrated in Fig. 8. This provides evidence that the conceptual idea of a population based measure, proposed in this paper, enhances the performance of a measure to separate networks from different categories.

On a note of caution, we would like to emphasize that the discriminating ability of the diversity score is not solely due to its population character, instead, it is due to the combination of its population character and the individual diversity score, , (see Eqn. 35), on which is based. From Fig. 8 one can learn about the influence of the sample size, but it does not give information about the influence of the individual diversity score. For this reason, we investigated the influence of the individual diversity score by altering its definition. For example, using only a subset of the four variables on which is based on (see Eqn. 30 to 33), we found that a population based version of such a measure does actually not lead to the discrimination of different network types. Hence, only the combination of an appropriate individual diversity score with a population approach results in the favorable characteristics of the diversity score.

In the section ‘Characterizing the complexity of networks’ we provided a characterization of complexity. The connection between this characterization, as given in Fig. 1, and our results in Fig. 8, is given by the cumulative distribution function (CDF) [81] of the densities in Fig. 8. Exemplarily, we show the CDF for and . Hence, the score (y-axis) in Fig. 1 can be identified with the cumulative distribution function of the probability density of the diversity score.

Finally, we show in Fig. 9 the influence of the sample size S on the mean individual diversity score , corresponding to , for networks of size . These results show that this mean value is largely constant for different values of the sample size S demonstrating that the unbiased estimator [83] given by Eqn. 37 provides good estimates in practice, even for small sample sizes. In addition, this figure demonstrates that very small sample sizes are not recommendable to use because the expected variability of the estimates is quite large.

Application to Real Networks

Finally, we apply the network diversity score to four real networks. We use two social networks representing coauthorship networks between scientists working in high-energy physics (hep, ) [84] and network science (net, ) [85], a technological network representing the Western States Power Grid of the United States (power, ) [82] and a biological network representing the protein-protein interactions in Helicobacter pylori (hpylo, ) [86], which is a bacterium that can be found in the stomach. The number in brackets refers to the number of nodes in the giant connected component of these networks, we use in the following for our analysis.

Because we have only one network for each of these four networks to which we can apply the network diversity score, we utilize the following property of complexity. It is generally assumed that one aspect of the complexity of an object is the presence of a hierarchical organization structure [10], [87], [88]. This implies that not only the whole object itself is complex but also a sufficiently large components of it. For our analysis, we utilize this by randomly selecting subnetworks from a network G. That means, we obtain a sample of S networks from one network by generating randomly subnetworks with n vertices from G. This way we obtain a sample of networks , whereas each network has been sampled from the network G, i.e.,(40)

that approximates a sample from an underlying network model. Practically, we generate the subnetworks by a random walk. Starting from an initial vertex that is randomly chosen from all vertices of the network , a subnetwork is defined by the first unique vertices visited by the random walk. This allows, first, to generate a sample of networks from a network model although only one network is available. Second, the size of each network can be set to a fixed value n. This allows the comparison of networks with a different size, because the size of the networks in the samples have all the same number of vertices.

In Fig. 10 we show the results for these four networks. In addition, we included results for random networks (red curve) generated with the Erdös-Réyni model. The x-axis gives the size of the subnetworks, n. The sample size for this analysis was and we averaged all results over 100 independent samples. That means for Fig. 10 we analyzed a total of networks. Overall, one can see that random networks lead to the lowest values of the density score and for subnetworks of size the distances between the individual networks are largely constant. This indicates that for the studied networks subnetworks of size are sufficiently large to capture the complexity of the whole networks.

thumbnail
Figure 10. Logarithm of the network diversity score in dependence on the size of the sampled subnetworks.

Every point on these curves is averaged over 100 independent samples of size S = 10.

https://doi.org/10.1371/journal.pone.0034523.g010

Discussion

In this paper we investigated the behavior of 17 network measures with respect to their ability to categorize the structural complexity of networks systematically. Our analysis demonstrates that constructing a network measure in a way that it averages over a sample of networks from a population, enhances its capabilities to categorize different types of networks significantly. From our numerical results follow that this averaging property of the diversity score is key in order to achieve a perfect separation of the three different network types, ordered, complex and random networks, we investigated in our analysis. The crucial point here is that this averaging property reduces the importance of finding the right network measure that quantifies exactly what is meant by the structural complexity of a network. Due to the fact that the right network complexity measure is not known, we defined the diversity score multiplicatively composed of four individual scores, each one assessing different structural properties of a network. Hence, the combination of a network diversity score, which does not focus on a single structural property of a network but on multiple ones, together with the averaging over a sample of networks from a population, leads to a network measure that appears to be well adopted to the proposed task. We would like to emphasize that there are other complexity measures that also include the underlying population in the definition of the measure [14], [16], [17], [21], [22], however, all of these complexity measures have only been studied in the context of symbol sequences.

On a theoretical note, the averaging over a sample of networks from a population does not only have a very beneficial influence on the numerical categorization of different types of networks, but removes also a conceptual ambiguity present in all measures that assess only individual networks with respect to their complexity. As discussed in the ‘Methods’ section, a random network model is capable of generating complex networks too. Hence, theoretically, it is possible to generate different types of networks with the random network model. This leads inevitably to a miscategorizations of networks. In contrast, the diversity score proposed in this paper reduces this ambiguity by a factor of , with S being the sample size.

The categorization of networks with respect to their structural complexity is not only interesting for theoretical, but also practical reasons. For example, in molecular biology it is generally assumed that molecular interactions between proteins and molecules generate the biological function of cells and give raise to the phenotypic appearance of organisms. Due to the fact that a graphical representation of such molecular interactions is given by gene networks, it has been suggested to compare these networks structurally in order to identify aberrations of molecular functions [89][91]. As an extension of the above approach it seems natural assessing the structural complexity of gene networks, e.g., of regulatory networks, to distinguish different stages of complex diseases, like cancer or cardiovascular disease, from each other. For example, gene expression data from DNA microarrays could be used to infer a regulatory network for each patient which belongs to a certain stage or a grade of a disease. Then such a disease grade can be considered as a category from which the patients and their respective networks are sampled. In this way, our network score can be applied to compare patients from different disease stages or grades with each other. Given the pace with which the data in molecular biology increase due to steady technological innovations, one can expect such data sets to be available within the near future. Other, potential areas of application are the categorization of financial networks [92][94] or neural networks [95], [96].

Acknowledgments

We would like to thank Ricardo de Matos Simoes, Shailesh Tripathi and John Quackenbush for fruitful discussions. For our numerical simulations we used R [97] and the package QuACN [98].

Author Contributions

Conceived and designed the experiments: FES MD. Performed the experiments: FES MD. Analyzed the data: FES MD. Contributed reagents/materials/analysis tools: FES MD. Wrote the paper: FES MD.

References

  1. 1. Bar-Yam Y (1997) Dynamics of Complex Systems. Perseus Books. Y. Bar-Yam1997Dynamics of Complex Systems.Perseus Books
  2. 2. Chaitin G (1966) On the length of programs for computing finite binary sequences. Journal of the ACM. pp. 547–569.G. Chaitin1966On the length of programs for computing finite binary sequences.Journal of the ACM547569
  3. 3. Kolmogorov AN (1965) Three approaches to the quantitative definition of ‘information’. Problems of Information Transmission 1: 1–7.AN Kolmogorov1965Three approaches to the quantitative definition of ‘information’.Problems of Information Transmission117
  4. 4. López-Ruiza R, Mancinib H, Calbet X (1995) A statistical measure of complexity. Physics Letters A 209: 321–326.R. López-RuizaH. MancinibX. Calbet1995A statistical measure of complexity.Physics Letters A209321326
  5. 5. Nicolis G, Prigogine I (1989) Exploring Complexity. Freeman. G. NicolisI. Prigogine1989Exploring Complexity.Freeman
  6. 6. Prokopenko M, Boschetti F, Ryan A (2009) An information-theoretic primer on complexity, selforganization, and emergence. Complexity 15: 11–28.M. ProkopenkoF. BoschettiA. Ryan2009An information-theoretic primer on complexity, selforganization, and emergence.Complexity151128
  7. 7. Schuster H (2002) Complex Adaptive Systems. Scator Verlag. H. Schuster2002Complex Adaptive Systems.Scator Verlag
  8. 8. Solomonoff R (1960) A preliminary report on a general theory of inductive inference. Technical Report V-131, Zator Co., Cambridge, Ma. R. Solomonoff1960A preliminary report on a general theory of inductive inference.Technical Report V-131, Zator Co., Cambridge, Ma
  9. 9. Wolfram S (1983) Statistical mechanics of cellular automata. Phys Rev E 55: 601–644.S. Wolfram1983Statistical mechanics of cellular automata.Phys Rev E55601644
  10. 10. Badii R, Politi A (1997) Complexity: Hierarchical Structures and Scaling in Physics. Cambridge University Press, Cambridge. R. BadiiA. Politi1997Complexity: Hierarchical Structures and Scaling in Physics.Cambridge University Press, Cambridge
  11. 11. Bennett C (1988) Logical depth and physical complexity. Herken R, editor, The Universal Turing Machine– a Half-Century Survey, Oxford University Press. pp. 227–257.C. Bennett1988Logical depth and physical complexity.Herken R, editor, The Universal Turing Machine– a Half-Century Survey, Oxford University Press227257
  12. 12. Crutchfield JP, Young K (1989) Inferring statistical complexity. Phys Rev Lett 63: 105–108.JP CrutchfieldK. Young1989Inferring statistical complexity.Phys Rev Lett63105108
  13. 13. Emmert-Streib F (2010) Exploratory analysis of spatiotemporal patterns of cellular automata by clustering compressibility. Physical Review E 81: 026103.F. Emmert-Streib2010Exploratory analysis of spatiotemporal patterns of cellular automata by clustering compressibility.Physical Review E81026103
  14. 14. Emmert-Streib F (2010) Statistic Complexity: Combining Kolmogorov Complexity with an Ensemble Approach. PLoS ONE 5: e12256.F. Emmert-Streib2010Statistic Complexity: Combining Kolmogorov Complexity with an Ensemble Approach.PLoS ONE5e12256
  15. 15. Gell-Mann M, Lloyd S (1998) Information measures, effective complexity, and total information. Complexity 2: 44–52.M. Gell-MannS. Lloyd1998Information measures, effective complexity, and total information.Complexity24452
  16. 16. Grassberger P (1986) Toward a quantitative theory of self-generated complexity. Int J Theor Phys 25: 907–938.P. Grassberger1986Toward a quantitative theory of self-generated complexity.Int J Theor Phys25907938
  17. 17. Lloyd S, Pagels H (1988) Complexity as thermodynamic depth. Annals of Physics 188: 186–213.S. LloydH. Pagels1988Complexity as thermodynamic depth.Annals of Physics188186213
  18. 18. Zurek W, editor. (1990) Complexity, Entropy and the Physics of Information. Addison-Wesley, Redwood City. W. Zurek1990Complexity, Entropy and the Physics of Information.Addison-Wesley, Redwood City
  19. 19. Grassberger P (1989) Problems in quantifying self-generated complexity. Helvetica Physica Acta 62: 489–511.P. Grassberger1989Problems in quantifying self-generated complexity.Helvetica Physica Acta62489511
  20. 20. Li M, Vitányi P (1997) An Introduction to Kolmogorov Complexity and Its Applications. Springer. M. LiP. Vitányi1997An Introduction to Kolmogorov Complexity and Its Applications.Springer
  21. 21. Crutchfield J, Packard N (1983) Symbolic dynamics of noisy chaos. Physica D 7: 201–223.J. CrutchfieldN. Packard1983Symbolic dynamics of noisy chaos.Physica D7201223
  22. 22. Bialek W, Nemenman I, Tishby N (2001) Predictability, complexity, and learning. Neural Computation 13: 2409–2463.W. BialekI. NemenmanN. Tishby2001Predictability, complexity, and learning.Neural Computation1324092463
  23. 23. Bonchev D (1983) Information Theoretic Indices for Characterization of Chemical Structures. Research Studies Press, Chichester. D. Bonchev1983Information Theoretic Indices for Characterization of Chemical Structures.Research Studies Press, Chichester
  24. 24. Bonchev D, Rouvray DH (2005) Complexity in Chemistry, Biology, and Ecology. Mathematical and Computational Chemistry. Springer. New York, NY, USA. D. BonchevDH Rouvray2005Complexity in Chemistry, Biology, and Ecology.Mathematical and Computational Chemistry. Springer. New York, NY, USA
  25. 25. Janežić D, Miležević A, Nikolić S, Trinajstić N (2009) Topological complexity of molecules. Meyers R, editor, Encyclopedia of Complexity and System Science, Springer, volume 5. pp. 9210–9224.D. JanežićA. MileževićS. NikolićN. Trinajstić2009Topological complexity of molecules.Meyers R, editor, Encyclopedia of Complexity and System Science, Springer, volume 592109224
  26. 26. Bertz SH (1983) On the complexity of graphs and molecules. Bull Math Biol 45: 849–855.SH Bertz1983On the complexity of graphs and molecules.Bull Math Biol45849855
  27. 27. Bonchev D, Polansky OE (1987) On the topological complexity of chemical systems. King RB, Rouvray DH, editors, Graph Theory and Topology, Elsevier. pp 125–158. Amsterdam, The Netherlands. D. BonchevOE Polansky1987On the topological complexity of chemical systems.King RB, Rouvray DH, editors, Graph Theory and Topology, Elsevierpp 125–158. Amsterdam, The Netherlands
  28. 28. Kim J, Wilhelm T (2008) What is a complex graph? Physica A 387: 2637–2652.J. KimT. Wilhelm2008What is a complex graph?Physica A38726372652
  29. 29. Dancoff SM, Quastler H (1953) Information content and error rate of living things. Quastler H, editor, Essays on the Use of Information Theory in Biology, University of Illinois Press. pp. 263–274.SM DancoffH. Quastler1953Information content and error rate of living things.Quastler H, editor, Essays on the Use of Information Theory in Biology, University of Illinois Press263274
  30. 30. Linshitz H (1953) The information content of a battery cell. Quastler H, editor, Essays on the Use of Information Theory in Biology, University of Illinois Press. Urbana, IL, USA. H. Linshitz1953The information content of a battery cell.Quastler H, editor, Essays on the Use of Information Theory in Biology, University of Illinois PressUrbana, IL, USA
  31. 31. Morowitz H (1953) Some order-disorder considerations in living systems. Bull Math Biophys 17: 81–86.H. Morowitz1953Some order-disorder considerations in living systems.Bull Math Biophys178186
  32. 32. Minoli D (1975) Combinatorial graph complexity. Atti Accad Naz Lincei, VIII Ser, Rend, Cl Sci Fis Mat Nat 59: 651–661.D. Minoli1975Combinatorial graph complexity.Atti Accad Naz Lincei, VIII Ser, Rend, Cl Sci Fis Mat Nat59651661
  33. 33. Bonchev D (2003) Complexity in Chemistry. Introduction and Fundamentals. Taylor and Francis. Boca Raton, FL, USA. D. Bonchev2003Complexity in Chemistry.Introduction and Fundamentals. Taylor and Francis. Boca Raton, FL, USA
  34. 34. Constantine G (1990) Graph complexity and the laplacian matrix in blocked experiments. Linear and Multilinear Algebra 28: 49–56.G. Constantine1990Graph complexity and the laplacian matrix in blocked experiments.Linear and Multilinear Algebra284956
  35. 35. Jukna S (2006) On graph complexity. Comb Probab Comput 15: 855–876.S. Jukna2006On graph complexity.Comb Probab Comput15855876
  36. 36. Li M, Vitányi P (1997) An Introduction to Kolmogorov Complexity and Its Applications. Springer. M. LiP. Vitányi1997An Introduction to Kolmogorov Complexity and Its Applications.Springer
  37. 37. Bonchev D (1995) Kolmogorov’s information, shannon’s entropy, and topological complexity of molecules. Bulg Chem Commun 28: 567–582.D. Bonchev1995Kolmogorov’s information, shannon’s entropy, and topological complexity of molecules.Bulg Chem Commun28567582
  38. 38. Dehmer M, Mowshowitz A (2011) A history of graph entropy measures. Information Sciences 1: 57–78.M. DehmerA. Mowshowitz2011A history of graph entropy measures.Information Sciences15778
  39. 39. Dehmer M (2008) Information processing in complex networks: Graph entropy and information functionals. Appl Math Comput 201: 82–94.M. Dehmer2008Information processing in complex networks: Graph entropy and information functionals.Appl Math Comput2018294
  40. 40. Skorobogatov VA, Dobrynin AA (1988) Metrical analysis of graphs. Commun Math Comp Chem 23: 105–155.VA SkorobogatovAA Dobrynin1988Metrical analysis of graphs.Commun Math Comp Chem23105155
  41. 41. Diudea MV, Gutman I, J¨antschi L (2001) Molecular Topology. Nova Publishing. New York, NY, USA. MV DiudeaI. GutmanL. J¨antschi2001Molecular Topology.Nova Publishing. New York, NY, USA
  42. 42. Todeschini R, Consonni V, Mannhold R (2002) Handbook of Molecular Descriptors. Weinheim: Wiley-VCH. R. TodeschiniV. ConsonniR. Mannhold2002Handbook of Molecular Descriptors.WeinheimWiley-VCH
  43. 43. Adami C (2002) What is complexity? BioEssays 24: 1085–1094.C. Adami2002What is complexity?BioEssays2410851094
  44. 44. Langton C (1990) Computation at the edge of choas: phase transitions and emergent computation. Physica D 42: 12–37.C. Langton1990Computation at the edge of choas: phase transitions and emergent computation.Physica D421237
  45. 45. Ribeiro AS, Kauffman SA, Lloyd-Price J, Samuelsson B, Socolar JES (2008) Mutual information in random boolean models of regulatory networks. Phys Rev E 77: 011901.AS RibeiroSA KauffmanJ. Lloyd-PriceB. SamuelssonJES Socolar2008Mutual information in random boolean models of regulatory networks.Phys Rev E77011901
  46. 46. Newman M (2010) Networks: An Introduction. Oxford: Oxford University Press. M. Newman2010Networks: An Introduction.OxfordOxford University Press
  47. 47. Balaban AT (1982) Highly discriminating distance-based topological index. ChemPhysLett 89: 399–404.AT Balaban1982Highly discriminating distance-based topological index.ChemPhysLett89399404
  48. 48. Bonchev D, Trinajstić N (1977) Information theory, distance matrix and molecular branching. J Chem Phys 67: 4517–4533.D. BonchevN. Trinajstić1977Information theory, distance matrix and molecular branching.J Chem Phys6745174533
  49. 49. Latora V, Marchiori M (2001) Efficient behavior of small-world networks. Phys Rev Lett 87: 198701.V. LatoraM. Marchiori2001Efficient behavior of small-world networks.Phys Rev Lett87198701
  50. 50. Latora V, MarchioriM (2003) Economic small-world behavior in weighted networks. The European Physical Journal B Condensed Matter 32: 249–263.V. LatoraMarchioriM2003Economic small-world behavior in weighted networks.The European Physical Journal B Condensed Matter32249263
  51. 51. Gutman I (1991) Polynomials in graph theory. Bonchev D, Rouvray DH, editors, Chemical Graph Theory. Introduction and Fundamentals, Abacus Press. pp133–176. New York, NY, USA. I. Gutman1991Polynomials in graph theory.Bonchev D, Rouvray DH, editors, Chemical Graph TheoryIntroduction and Fundamentals, Abacus Press. pp133–176. New York, NY, USA
  52. 52. Dehmer M, Emmert-Streib F, Tsoy Y, Varmuza K (2011) Quantifying structural complexity of graphs: Information measures in mathematical chemistry. Putz M, editor, Quantum Frontiers of Atoms and Molecules, Nova Publishing. pp. 479–498.M. DehmerF. Emmert-StreibY. TsoyK. Varmuza2011Quantifying structural complexity of graphs: Information measures in mathematical chemistry.Putz M, editor, Quantum Frontiers of Atoms and Molecules, Nova Publishing479498
  53. 53. Gutman I, Zhou B (2006) Laplacian energy of a graph. Linear Algebra and its Applications 414: 29–37.I. GutmanB. Zhou2006Laplacian energy of a graph.Linear Algebra and its Applications4142937
  54. 54. Claussen JC (2007) Characterization of networks by the offdiagonal complexity. Physica A 365–373: 321–354.JC Claussen2007Characterization of networks by the offdiagonal complexity.Physica A365–373321354
  55. 55. Randić M (1975) On characterization of molecular branching. J Amer Chem Soc 97: 6609–6615.M. Randić1975On characterization of molecular branching.J Amer Chem Soc9766096615
  56. 56. Mowshowitz A (1968) Entropy and the complexity of the graphs I: An index of the relative complexity of a graph. Bull Math Biophys 30: 175–204.A. Mowshowitz1968Entropy and the complexity of the graphs I: An index of the relative complexity of a graph.Bull Math Biophys30175204
  57. 57. Wiener H (1947) Structural determination of paraffin boiling points. Journal of the American Chemical Society 69: 17–20.H. Wiener1947Structural determination of paraffin boiling points.Journal of the American Chemical Society691720
  58. 58. Rashevsky N (1955) Life, information theory, and topology. Bull Math Biophys 17: 229–235.N. Rashevsky1955Life, information theory, and topology.Bull Math Biophys17229235
  59. 59. Trucco E (1956) A note on the information content of graphs. Bull Math Biol 18: 129–135.E. Trucco1956A note on the information content of graphs.Bull Math Biol18129135
  60. 60. Mowshowitz A (1968) Entropy and the complexity of graphs II: The information content of digraphs and infinite graphs. Bull Math Biophys 30: 225–240.A. Mowshowitz1968Entropy and the complexity of graphs II: The information content of digraphs and infinite graphs.Bull Math Biophys30225240
  61. 61. Mowshowitz A (1968) Entropy and the complexity of graphs III: Graphs with prescribed information content. Bull Math Biophys 30: 387–414.A. Mowshowitz1968Entropy and the complexity of graphs III: Graphs with prescribed information content.Bull Math Biophys30387414
  62. 62. Cormen TH, Leiserson CE, Rivest RL (1990) Introduction to Algorithms. MIT Press. TH CormenCE LeisersonRL Rivest1990Introduction to Algorithms.MIT Press
  63. 63. Dijkstra E (1959) A note on two problems in connection with graphs. Numerische Math 1: 269–271.E. Dijkstra1959A note on two problems in connection with graphs.Numerische Math1269271
  64. 64. Balaban AT, Balaban TS (1991) New vertex invariants and topological indices of chemical graphs based on information on distances. J Math Chem 8: 383–397.AT BalabanTS Balaban1991New vertex invariants and topological indices of chemical graphs based on information on distances.J Math Chem8383397
  65. 65. Buckley F, Harary F (1990) Distance in Graphs. Addison Wesley Publishing Company. F. BuckleyF. Harary1990Distance in Graphs.Addison Wesley Publishing Company
  66. 66. Li X, Gutman I (2006) Mathematical Aspects of Randić-Type Molecular Structure Descriptors. Mathematical Chemistry Monographs. University of Kragujevac and Faculty of Science Kragujevac. X. LiI. Gutman2006Mathematical Aspects of Randić-Type Molecular Structure Descriptors.Mathematical Chemistry Monographs. University of Kragujevac and Faculty of Science Kragujevac
  67. 67. Dehmer M, Müller L, Graber A (2010) New polynomial-based molecular descriptors with low degeneracy. PLoS ONE 5: M. DehmerL. MüllerA. Graber2010New polynomial-based molecular descriptors with low degeneracy.PLoS ONE5
  68. 68. Garey MR, Johnson DS (1979) Computers and Intractability: A Guide to the Theory of NPCompleteness. Series of Books in the Mathematical Sciences. W. H. Freeman. MR GareyDS Johnson1979Computers and Intractability: A Guide to the Theory of NPCompleteness.Series of Books in the Mathematical Sciences. W. H. Freeman
  69. 69. Mowshowitz A, Mitsou V (2009) Entropy, orbits and spectra of graphs. Dehmer M, Emmert-Streib F, editors, Analysis of Complex Networks: From Biology to Linguistics, Wiley-VCH. pp. 1–22.A. MowshowitzV. Mitsou2009Entropy, orbits and spectra of graphs.Dehmer M, Emmert-Streib F, editors, Analysis of Complex Networks: From Biology to Linguistics, Wiley-VCH122
  70. 70. Devillers J, Balaban AT (1999) Topological Indices and Related Descriptors in QSAR and QSPR. Gordon and Breach Science Publishers. Amsterdam, The Netherlands. J. DevillersAT Balaban1999Topological Indices and Related Descriptors in QSAR and QSPR.Gordon and Breach Science Publishers. Amsterdam, The Netherlands
  71. 71. Sagan H (1989) Boundary and Eigenvalue Problems in Mathematical Physics. Dover Publications. H. Sagan1989Boundary and Eigenvalue Problems in Mathematical Physics.Dover Publications
  72. 72. Pons P, LatapyM (2005) Computing communities in large networks using random walks. Yolum p, Güngör T, Gürgen F, Özturan C, editors, Computer and Information Sciences - ISCIS 2005, Springer Berlin/Heidelberg, volume 3733 of Lecture Notes in Computer Science. pp. 284–293.P. PonsLatapyM2005Computing communities in large networks using random walks.Yolum p, Güngör T, Gürgen F, Özturan C, editors, Computer and Information Sciences - ISCIS 2005, Springer Berlin/Heidelberg, volume 3733 of Lecture Notes in Computer Science284293
  73. 73. Van Dongen S (2000) Graph clustering by flow simulation. Ph.D. thesis, Centers for mathematics and computer science (CWI), University of Utrecht. S. Van Dongen2000Graph clustering by flow simulation.Ph.D. thesis, Centers for mathematics and computer science (CWI), University of Utrecht
  74. 74. Ziv E, Middendorf M, Wiggins CH (2005) Information-theoretic approach to network modularity. Phys Rev E 71: 046117.E. ZivM. MiddendorfCH Wiggins2005Information-theoretic approach to network modularity.Phys Rev E71046117
  75. 75. Chung FRK (1997) Spectral Graph Theory. American Mathematical Society. FRK Chung1997Spectral Graph Theory.American Mathematical Society
  76. 76. Milo R, Shen-Orr S, Itzkovitz S, Kashtan N, Chklovskii D, et al. (2002) Network motifs: simple building blocks of complex networks. Science 298: 824–7.R. MiloS. Shen-OrrS. ItzkovitzN. KashtanD. Chklovskii2002Network motifs: simple building blocks of complex networks.Science2988247
  77. 77. Erdös P, Rényi A (1959) On random graphs. I. Publicationes Mathematicae 6: 290–297.P. ErdösA. Rényi1959On random graphs.I. Publicationes Mathematicae6290297
  78. 78. Gilbert EN (1959) Random graphs. Annals of Mathematical Statistics 20: 1141–1144.EN Gilbert1959Random graphs.Annals of Mathematical Statistics2011411144
  79. 79. Albert R, Barabasi A (2002) Statistical mechanics of complex networks. Rev of Modern Physics 74: 47.R. AlbertA. Barabasi2002Statistical mechanics of complex networks.Rev of Modern Physics7447
  80. 80. Barabási AL, Albert R (1999) Emergence of scaling in random networks. Science 206: 509–512.AL BarabásiR. Albert1999Emergence of scaling in random networks.Science206509512
  81. 81. Feller W (1968) An Introduction to Probability Theory and Its Applications Vol. 1. John Wiley & Sons. W. Feller1968An Introduction to Probability Theory and Its Applications Vol.1. John Wiley & Sons
  82. 82. Watts D, Strogatz S (1998) Collective dynamics of ‘small-world’ networks. Nature 393: 440–442.D. WattsS. Strogatz1998Collective dynamics of ‘small-world’ networks.Nature393440442
  83. 83. Lehman E, Casella G (1999) Theory of point estimation. New York: Springer. E. LehmanG. Casella1999Theory of point estimation.New YorkSpringer
  84. 84. NewmanMEJ (2001) The structure of scientific collaboration networks. Proceedings of the National Academy of Sciences of the United States of America 98: 404–409.NewmanMEJ2001The structure of scientific collaboration networks.Proceedings of the National Academy of Sciences of the United States of America98404409
  85. 85. Newman MEJ (2006) Finding community structure in networks using the eigenvectors of matrices. Phys Rev E 74: 036104.MEJ Newman2006Finding community structure in networks using the eigenvectors of matrices.Phys Rev E74036104
  86. 86. Xenarios I, Rice DW, Salwinski L, Baron MK, Marcotte EM, et al. (2000) DIP: the Database of Interacting Proteins. Nucl Acids Res 28: 289–291.I. XenariosDW RiceL. SalwinskiMK BaronEM Marcotte2000DIP: the Database of Interacting Proteins.Nucl Acids Res28289291
  87. 87. Ceccatto HA, Huberman BA (1988) The complexity of hierarchical systems. Physica Scripta 37: 145.HA CeccattoBA Huberman1988The complexity of hierarchical systems.Physica Scripta37145
  88. 88. Ravasz E, Somera AL, Mongru DA, Oltvai ZN, Barabasi AL (2002) Hierarchical organization of modularity in metabolic networks. Science 297: 1551–1555.E. RavaszAL SomeraDA MongruZN OltvaiAL Barabasi2002Hierarchical organization of modularity in metabolic networks.Science29715511555
  89. 89. Emmert-Streib F (2007) The chronic fatigue syndrome: A comparative pathway analysis. Journal of Computational Biology 14: 961–972.F. Emmert-Streib2007The chronic fatigue syndrome: A comparative pathway analysis.Journal of Computational Biology14961972
  90. 90. Emmert-Streib F, Glazko G (2011) Pathway analysis of expression data: deciphering functional building blocks of complex diseases. PLoS Computational Biology 7: e1002053.F. Emmert-StreibG. Glazko2011Pathway analysis of expression data: deciphering functional building blocks of complex diseases.PLoS Computational Biology7e1002053
  91. 91. Schadt E (2009) Molecular networks as sensors and drivers of common human diseases. Nature 461: 218–223.E. Schadt2009Molecular networks as sensors and drivers of common human diseases.Nature461218223
  92. 92. Boginski V, Butenko S, Pardalos P (2005) Statistical analysis of financial networks. Computational Statistics and Data Analysis 48: 431–443.V. BoginskiS. ButenkoP. Pardalos2005Statistical analysis of financial networks.Computational Statistics and Data Analysis48431443
  93. 93. Emmert-Streib F, Dehmer M (2010) Identifying Critical Financial Networks of the DJIA: Towards a Network-based Index. Complexity 16: 24–33.F. Emmert-StreibM. Dehmer2010Identifying Critical Financial Networks of the DJIA: Towards a Network-based Index.Complexity162433
  94. 94. Emmert-Streib F, Dehmer M (2010) Influence of the Time Scale on the Construction of Financial Networks. PLoS ONE 5: e12884.F. Emmert-StreibM. Dehmer2010Influence of the Time Scale on the Construction of Financial Networks.PLoS ONE5e12884
  95. 95. Kaiser M, Hilgetag CC, Kötter R (2010) Hierarchy and dynamics of neural networks. Frontiers in Neuroinformatics 4: M. KaiserCC HilgetagR. Kötter2010Hierarchy and dynamics of neural networks.Frontiers in Neuroinformatics4
  96. 96. Sporns O (2011) Networks of the brain. Cambridge, Mass: MIT Press. O. Sporns2011Networks of the brain.Cambridge, MassMIT Press
  97. 97. R Development Core Team (2008) R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria.R Development Core Team2008R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria.ISBN 3-900051-07-0. ISBN 3-900051-07-0.
  98. 98. Mueller LA, Kugler KG, Dander A, Graber A, Dehmer M (2010) QuACN - An R Package for Analyzing Complex Biological Networks Quantitatively. Bioinformatics. LA MuellerKG KuglerA. DanderA. GraberM. Dehmer2010QuACN - An R Package for Analyzing Complex Biological Networks Quantitatively.Bioinformatics