Physica A 456 (2016) 167–175

Contents lists available at ScienceDirect

Physica A
journal homepage: www.elsevier.com/locate/physa

A geometric graph model for citation networks of

exponentially growing scientific papers
Zheng Xie a,b,∗ , Zhenzheng Ouyang a , Qi Liu a , Jianping Li a
a
College of Science, National University of Defense Technology, Changsha, 410073, China
b
Centre for Networks and Collective Behaviour, Department of Mathematics, University of Bath, Bath, BA2 7AY, UK

highlights
• A geometric graph is proposed to model the citation networks of exponentially growing papers.
• The model expresses certain factors engendering citations, e.g. the relativity of contents.
• The model predicts certain features of the citation networks, e.g. in-degree assortativity.

article info abstract

Article history: In citation networks, the content relativity of papers is a precondition of engendering
Received 22 August 2015 citations, which is hard to model by a topological graph. A geometric graph is proposed to
Received in revised form 6 January 2016 predict some features of the citation networks with exponentially growing papers, which
Available online 1 April 2016
addresses the precondition by using coordinates of nodes to model the research contents
of papers, and geometric distances between nodes to diversities of research contents
Keywords:
between papers. Citations between modeled papers are drawn according to a geometric
Geometric graph
Citation network
rule, which addresses the precondition as well as some other factors engendering citations,
Assortativity namely academic influences of papers, aging of those influences, and incomplete copying of
Modelling references. Instead of cumulative advantage of degree, the model illustrates that the scale-
Bibliometric free property of modeled networks arises from the inhomogeneous academic influences of
Causal network modeled papers. The model can also reproduce some other statistical features of citation
networks, e.g. in- and out-assortativities, which show the model provides a suitable tool to
understand some aspects of citation networks by geometry.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction

Citation networks constructed from scientific papers are important research objects of scientometrics, in which each
node represents a paper, and each edge represents a citation of one paper by another. Modeling these networks provides a
window on understanding hot topics of research, the emergence and propagation of academic thoughts in scientific society,
etc. [1–4]. Many of the empirically observed citation networks are found to be scale-free (their in-degree distributions
have a power-law tail), clustering, and assortative (in the sense of in- and out-degrees respectively). Seeking mechanisms
to illustrate one or more of those properties has attracted extensive attention [5,6].
There have existed several important studies of the scale-free property of citation networks. Price noted the ‘‘cumulative
advantage’’ of citation behavior: highly cited scientific papers accumulate additional citations more quickly than papers that

∗ Corresponding author at: College of Science, National University of Defense Technology, Changsha, 410073, China.
E-mail address: [email protected] (Z. Xie).

http://dx.doi.org/10.1016/j.physa.2016.03.018
0378-4371/© 2016 Elsevier B.V. All rights reserved.
168 Z. Xie et al. / Physica A 456 (2016) 167–175

have fewer citations. He abstractly expressed this phenomenon by a rule: the probability that a paper receives a citation is
proportional to the number of citations it has received, which successfully predicts the scale-free property [7–9]. In network
science, cumulative advantage is also called preferential attachment. Price model has been generalized to illustrate other
properties of citation networks in various contexts [10–13], e.g. Goldberg et al. set the number of citations given by new
papers to be random variables drawn from a lognormal distribution, which fits the out-degree distributions of empirical
data well (the model A in Ref. [14]).
It is empirically observed that the probability for a paper to get cited decreases as its age increases, which is called aging
phenomenon of citation behavior. Some models introduced time decay to the cumulative advantage, namely the probability
of an existing paper to be cited is proportional to its current in-degree multiplied by a decay factor dependent on its age
[15,16], e.g. exponential decay factor (the model B in Ref. [14]). Aging makes citation bursts typically occur in the early life
of a paper. Eom et al. generalized the Price model to simulate the bursts, in which a new paper cites a constant number of
existing papers by a linear preferential attachment with time dependent initial attractiveness [17].
Empirical data have a positive clustering coefficient, which is zero in theory for the networks generated by
aforementioned models. Krapivsky et al. noted that the authors of a new paper may not only cite a paper, but also could
cite some references of the cited paper. They called this phenomenon copying, and mimicked it by a rule: a new node
connects to a randomly selected node, as well as all the ancestors of the selected node, which successfully predicts the
scale-free property of citation networks, and clustering as well [18]. In reality, copying is incomplete: a paper unlikely cites
all references of the papers it cited. In the misprints propagation model [19] and the model C in Ref. [14], incomplete copying
is realized by a one-step random walk from a cited paper. Incomplete copying can also be added to the cumulative advantage
with time decay to address the scale-free, clustering and aging simultaneously [20].
The exponents of in-degree distributions of citation networks vary from data to data, but the predicted power-law
exponents (if provided) given by all above models are fixed. Peterson et al. proposed a model to address this problem,
which involves a direct mechanism: a new paper cites an old paper randomly; an indirect mechanism, which is a kind of
incomplete copying [21]. The model can generate networks with similar full in-degree distributions of empirical citation
networks (the exponents of in-degree distributions of which can be tuned) namely not only similar on the power-law tails
but also on the foreparts. The indirect mechanism makes the modeled networks have a positive clustering coefficient as
well.
The theory of random geometric graphs (RGGs) enables research into networks via geometry [22–26]. The nodes of some
RGGs are points chosen at random in the space time, e.g. through a Poisson point process, and they are connected by edges
if they are causally connected [27]. The causal relationship is induced by light cones in the space-time. Meanwhile, the idea
of a paper is inspired by its references at certain levels, so citation behavior could be regarded as a causal relationship. We
have proposed a RGG built on a cluster of concentric circles (so called it CC model) for citation networks [28], in which
the influences of modeled papers are expressed by geometric zones liking light cones, and a paper i cites a paper j if the
influential zone of i contains j. The model can capture the scale-free and clustering properties of empirical networks, but
has a range of shortcomings, e.g. the out-degree distributions of empirical data cannot be well simulated.
Besides the causal property of citations, using RGGs can also illustrate an important precondition of engendering citations,
namely the content relativity of papers, by spatial coordinates of nodes: diversities of research contents between papers
are illustrated by geometric distances between nodes. Here, we continue to use RGG built on the space time of the CC
model to predict certain statistical features of the citation networks with exponentially growing papers, and address
some shortcomings of the CC model as well. We seek a geometric mechanism to simultaneously express certain factors
engendering citations, namely academic influences, the aging of those influences, relativity of contents, and incomplete
copying of references. The model shows, besides the cumulative advantage, the scale-free property of citation networks
can also be explained as a consequence of the inhomogeneous academic influences of scientific papers, through which
some papers gain more citations because they are likely to have wider influences than others. We also examined how the
model predicts some other statistical features of empirical networks, namely the out-degree distribution, the scaling relation
between local clustering coefficients and in-degrees, assortativities for in- and out-degrees.
This report is organized as follows. The model is described in Section 2. The degree distributions, clustering and
assortativity of the modeled networks are analyzed in Sections 3–5 respectively. The conclusion is drawn in Section 6.

2. The model

Normally, empirical citation networks of scientific papers are directed acyclic graphs (DAGs): only newer papers can
cite older papers. However, some preprinting papers would cite each other, which happens rarely. The geometric DAG
proposed here consists of ‘‘papers’’ (nodes) and ‘‘citations’’ (edges) between those papers. In some citation networks, the
annual numbers of papers grow exponentially, e.g. the citation network DBLP 2013-09-29 (Table 1) collected by Tang et al.
for the papers in DBLP dataset, which are published in the period from 1936-01-01 to 2013-09-29 [29] (Fig. 1(a)). We focus
on simulating the exponentially growing case and compare some properties of a network generated by our model to those
of DBLP 2013-09-29.
In our model, the nodes are sprinkled on a cluster of concentric circles in a (2 + 1)-dimensional spacetime with
circumference polar coordinates {r , θ , t } (Fig. 2). The angular coordinates of nodes could be regarded as research contents
of papers. So diversities of research contents between papers could be abstractly expressed by geometric distances between
 Z. Xie et al. / Physica A 456 (2016) 167–175 169

Table 1
Certain statistical indicators of the analyzed citation networks.
Network Nodes Edges CC In-AC Out-AC PG MO

DBLP 2013-09-29 2,084,055 2,244,018 0.070 0.036 0.093 0.973 0.771

Cit-HepTh 27,770 352,807 0.165 0.041 0.096 0.987 0.648
Cit-HepPh 34,546 421,578 0.148 0.077 0.111 0.995 0.725
Modeled network 123,008 171,891 0.390 0.120 0.290 0.452 0.992
The indicators are clustering coefficient (CC), in- and out-assortativity coefficients (In-AC, Out-AC), the node proportion of the giant component in the total
(PG), and modularity (MO). The first network comes from the papers published before 2013-09-29 in DBLP dataset. The next two networks come from the
papers (which are published in the period from 1993-01 to 2003-04) of arXiv in high energy physics phenomenology and in high energy physics theory
respectively [30]. The fourth network is generated by our model, the parameters of which are α = 0.8, β = 0.05, δ = 5, l = 0.02, m = 1, p = 1, q =
0.5, r = 0.002, T = 390 and f (k) ∼ k−2.5 , k ∈ [3, 60].

(a) DBLP 1936–2012. (b) Modeled network.

Fig. 1. Evolutionary trends of annual numbers of papers. Panel (a) shows the trend for the papers of DBLP published in 1936–2012, and Panel (b) for the
modeled network in Table 1. The coefficient of determination (R2 ) and root mean squared error (RMSE) are used to measure the goodness of fits.

Fig. 2. Illustration of the model. The large nodes illustrate highly cited papers, and their influential zones are represented by light blue areas. Different
sizes of academic influences are expressed by different zonal sizes. It demonstrates the core and complement mechanisms of citations respectively: if node
a5 belongs to the zones of nodes a1 and a3 , then it cites them; if node a6 is a multidisciplinary paper, it will not only cite the highly cited papers which have
zones cover a6 (e.g. a5 ) but also cite a number of existing papers (e.g. a4 ) to make its out-degree to be a random variable drawn from a given power-law
distribution. The numbers of citations received by newer nodes could be larger than those of older ones, e.g. the times cited of a4 are larger than that of a2 .

nodes. The time t is related to physical time, which can be explained as the tth unit of time, such as tth week, tth month,
etc. Some nodes are selected to attach to specific zones to express their academic influences, which are called ‘‘highly cited
papers’’ here. Meanwhile, some nodes are selected to be ‘‘multidisciplinary papers’’, the out-degree of which is drawn from
a power-law distribution.
Suppose a modeled network has m[elt ] papers published in the tth unit of time (t = 1, 2, . . . , T ), in which some are
highly cited papers and some are multidisciplinary papers, where l ∈ R+ , m ∈ Z+ , and [·] is the rounding function. For time
t = 1, 2, . . . , T ∈ Z+ , the modeled papers and citations are created by following steps.

Step 1. Generate a circle Ct = ([elt ], θ , t ), sprinkle m[elt ] nodes (which are regarded as papers) on Ct randomly and
uniformly (namely according to a Poisson point process), and fix nodes with their coordinates, e.g. i with (θi , ti );
170 Z. Xie et al. / Physica A 456 (2016) 167–175

Step 2. Select q percent new papers randomly as potential highly cited papers to attach to specific zones: the zone of a
such paper i is defined as an interval of angular coordinate with center θi and arc-length β(θi )e−α lti −(1−α)ltc , where
tc is the current time, α ∈ (0, 1), and β is a positive and piecewise-constant function;
Step 3. Each new paper j cites with a probability p to the potential highly cited papers that have zones covering j and are
published δ earlier than j, where δ ∈ Z+ ;
Step 4. Select r percent new papers randomly as multidisciplinary papers to continually cite a number of existing papers
randomly to make the lengths of references (out-degrees) of those papers to be random variables drawn from a
power-law distribution f (k).

The parameters p, q and r are properly chosen to make the number of citations generated in Step 3 to be the core while
that in Step 4 to be the complement. There are some intuitive explanations for the formula of zonal sizes. First, if two highly
cited papers discuss about the same content, the cumulative advantage would make the older one receive more citations
than the newer one, so the formula is reasonable to consider to be a decreasing function of t. Second, the number of citations
may differ in research contents due to their inhomogeneous attractiveness, so β(·) is introduced to the formula. Third, aging
of the influences of papers occurs such that the probability of a paper getting cited decreases with the growth of its age, so
it is reasonable to consider the formula to be a decreasing function of tc .
Our model is developed to address following points of particular interest to us, which are the reasons for why we
introduce so many parameters. The details about how our model gives a solution to those points are shown in following
sections. We discuss how some typical models deal with those points as well.
Firstly, almost existing models focus on power-law tails, e.g. the rules of linear preferential attachment and complete
copying predict power-law in-degree distributions with a fixed exponent. We are interested here in more features about
the in- and out-degree distributions of empirical data, namely the hook heads in the out-degree distributions and power-law
tails (the exponents of which vary from data to data) in both of the in- and out-degree distributions. Through calculations
and numerical simulations in Section 4, we reveal the modeled networks capture those features. Especially, the exponents
of the predicted in-degree distributions can be tuned by the parameter α in the formula of zonal sizes.
Secondly, the global clustering coefficient (GCC) is the fraction of connected triples of nodes which also form ‘‘triangles’’.
Copying citations is a sound explanation for why citation networks have a positive GCC. The values of GCC differ from an
empirical network to another (Table 1), which cannot be simulated by the linear attachment (too low), the complete copying
(too high), and the CC model (non-tunable). Aging of the influences of papers is a reasonable factor for why the GCCs of
empirical data are lower than those predicted by the complete copying, because old papers hardly receive citations. So
incomplete copying is more realistic, which is expressed by different ways in some models [20], e.g. the model C in Ref. [14].
The GCCs predicted by those models can be tuned by the proportion of citations generated through incomplete copying. Our
model gives a sympathetical expression of aging (expressed by the decreasing sizes of influential zones with the growth of
time) and incomplete copying (realized by the parameter p in Step 3), the detail of which is shown in Section 5.
Thirdly, the empirical citation networks are in- and out-assortative. Simulations show that the first property can be
reproduced by linear preferential attachment, and both by complete copying, the reasons of which could be inferred from
their mechanisms. The proposed model can also reproduce those properties, the reason of which is explained by using
analytical formulae in Section 6.
Particularly, except the non-tunable global clustering coefficient, the CC model has some other obvious shortcomings,
e.g. the out-degree distributions of some empirical data have a short power-law tail. Now the multidisciplinary papers make
the out-degree distributions of modeled networks also have a short power-law tail, namely the function f (k) in Step 4. In
reality, some newly published papers can also gain more citations than some older papers that discuss about the same
content, which cannot happen in the CC model. Due to the selection in Step 2, a paper published at the earliest times could
not be a highly cited paper, and a new paper could receive many citations if it is selected to be a highly cited one.
Simulations show modeled networks have giant components (in sense of weak, namely ignore the directions of edges)
and clear community structures as the empirical networks do (Table 1). In fact, the papers in the same zone are very likely
to belong to the same community, and the zones are loosely connected, or not connected at all (the level of connection can
be tuned by the parameter r and the range of f (k) in Step 4) by few multidisciplinary papers. This causes the edges within
communities to be significantly more than that between communities. Therefore, there is a reciprocal relationship between
the clearness of communities and the node proportion of the giant component in the total.
The fitting function of the annual number of papers in DBLP 2013-09-29 is N (t ) = 70e0.1052t (Fig. 1(a)). To make the
modeled time span match with, and the annual number of modeled papers proportional to those of the empirical data
respectively, we set the unit of the model parameter t to be l/0.1052 ≈ 0.2 year, which makes the modeled time span to
be Tl/0.1052 ≈ 78 years (Fig. 1(b)). Since our model generates networks by a global selection rule, it is hard to generate
a network with a size comparable to DBLP 2013-09-29, which needs m = 14. We set m = 1 for the modeled network in
Table 1, namely there is only one paper in the beginning (because [e0.02 ] = 1) which happens in many classical citation
network models, e.g. Price model. We further set other parameters to make the modeled average degree similar to that of
the empirical data, and assume that the papers published apart less than a year (namely δ = 5) cannot cite each other.
 Z. Xie et al. / Physica A 456 (2016) 167–175 171

(a) DBLP 2013-09-29. (b) Modeled network.

(c) DBLP 2013-09-29. (d) Modeled network.

functions are the generalized Poisson distribution p(k) = a(a + bk)k−1 e−a−bk /k! for the hooks in Panels
Fig. 3. Out- and in-degree distributions. The fitting
(a,b), and the power-law distribution p(k) = k−β / l=0 (l + xmin )−β for the tails in Panels (c,d) (which is fitted by the method in Ref. [32]).
∞

3. Out- and in-degree distributions

The out-degree distributions of empirical citation networks exhibit power-law tails and ‘‘hooks’’ in the foreparts [25]
(Fig. 3(a)). The hooks can be adequately fitted by generalized Poisson distributions, which hint at the potential that the
predominant citation behaviors are governed by generalized Poisson processes. Many events in real world (in which the
probability of occurrence of a single event is small and is affected by previous occurrences) come from generalized Poisson
processes [31]. Citation behavior can be viewed as a low-rate event (its realization selects few papers out of a great many
papers) and the probability of occurrence of a citation could be affected by previous occurrences (e.g. a citation generated
by copying). Now we drive the underlying formulae of the head and tail of the modeled out-degree distribution to show
how our model generates a similar hook and power-law tail (Fig. 3(b)).
We initially consider the out-degrees of the non-multidisciplinary papers, and then consider those of the
multidisciplinary papers. Suppose a paper i is a non-multidisciplinary paper. If i belongs to the influential zone of j, then
we have ∆(θi , θj ) < β(θj )e−l(α tj +(1−α ti )) and tj < ti − δ . When β(θj )e−l(α tj +(1−α ti )) is small enough, we have β(θj ) ≈ β(θi ),
because β(·) is a piecewise constant function. Hence the expected out-degree k+ (θi , ti ) of i is approximately equal to

β(θi )mpq e−(1−α)l − e−(1−α)l(ti −δ)

ti −δ
m  β(θi )pq
k+ (θi , ti ) = × [eltj ] ≈ , (1)
2π t =1 e(α tj +(1−α)ti )l 2π (1 − α)l
j

which is not a constant but increases with the growth of ti . It is observed that the annual average length of a paper’s references
is a monotone increasing sequence for some journals [4,28]. So the increasing property of Eq. (1) is reasonable. When ti ≫ 1,
we can take the expected out-degree k+ (θi , ti ) to be independent of ti and write k+ (θi ) instead
β(θi )mpq −(1−α)l
k+ (θi ) = e , (2)
2π (1 − α)l
which holds for the majority papers, because the annual number of papers grows exponentially. It is reasonable to consider
that the number of a paper’s references cannot grow to infinity and should have an upper bound. Hence, the result given by
Eq. (2) is also reasonable, because that a bounded monotonic sequence has a limit.
172 Z. Xie et al. / Physica A 456 (2016) 167–175

The out-degrees of papers will not be exactly equal to their expected values because the papers are distributed according
to a Poisson point process, and so need to be averaged with the Poisson distribution. Hence the out-degree distribution of
−m (k) is approximately to be a mixture Poisson distribution:
+
non-multidisciplinary papers Pnon
2π 2π
   
k+ (θi , ti )k e−k (θi ,ti ) k+ (θi )k e−k (θi )
 T +1  +  +
1 1
Pnon−m (k) =
+
dθi dti ≈ dθi . (3)
2π T 1 0 k! 2π 0 k!

The multidisciplinary papers make the out-degree distribution P + (k) have a power-law tail. More precisely, we have
P + (k) ≈ (1 − q)Pnon
+
−m (k) + qf (k), (4)
where q is the proportion of multidisciplinary papers, and f (k) is a power-law distribution (both are defined in Step 4 of
the model description). As Eq. (4) shows, the model cannot exactly reproduce the hooks of the out-degree distributions of
empirical networks, which is a generalized Poisson distribution, but not a mixture Poisson. Overcoming this shortcoming is
indicative of the need for further research.
The tails of in-degree distributions of empirical citation networks are well fitted by power-law distributions (Fig. 3(c)),
which has been considered as a consequence of cumulative advantage of degrees or complete copying. In our model, the
highly cited papers (which have an influential zone) are responsible for the power-law tails of the modeled in-degree
distributions (Fig. 3(d)). This hints that the power-law tails may also alternatively be interpreted as a consequence of the
inhomogeneous sizes of academic influences.
The expected in-degree k− (θj , tj ) of a highly cited paper j is the probability p multiplied by the expected number of
papers belonging to its zone, namely
T
m  β(θj )p β(θj )mp α(T −tj )l
k− (θj , tj ) ≈ (α tj +(1−α)s)l
× [els ] ≈ e . (5)
2π s=t +δ e 2π α l
j

The first approximation is due to the ignorance of the in-degrees contributed by the complement mechanism Step 4, because
the order of magnitude of q is set to be smaller than that of p. The second approximation holds for T ≫ 1. The calculation of
underlying formula for the in-degree distribution in the large-k region (namely the tail) PL− (k) is the same as that in Ref. [26]
and is briefly listed as follows:
2π T +1
  
1 1
k− (θj , tj )k e−k (θj ,tj ) dtj
−
PL− (k) = dθj ∝ , (6)
2π k! 0 1 k
1
1+ α

where k ≫ 0 is needed for the approximation.

The multidisciplinary papers can slightly affect the forepart of in-degree distribution, because those papers cite some
existing papers randomly, which makes the forepart have a small component of Poisson distribution. Hence, the forepart of
in-degree distribution of the modeled network slightly deviates from the power-law distribution.

4. Local clustering coefficients and in-degrees

The local clustering coefficient (LCC) of a node is the connecting probability of two neighbors of the node. In empirical
citation networks, the large in-degree papers have low LCCs. A sound explanation is that some highly cited papers have
influences in many topics, and the papers in different topics may not cite each other even if they cite a common paper.
Moreover, considering the average LCC of in-degree k nodes C (k), the tails of C (k) are roughly proportional to 1/k. We
illustrate this property for the DBLP 2013-09-29 (Fig. 4(a)). The other two empirical networks also have this property, which
is not illustrated here. To show how the model generates a similar tail (Fig. 4(b)), we next derive the underlying formula of
the tails of C (k) for modeled networks.
Suppose i is a highly cited paper and ti is small enough. The large in-degree nodes of modeled networks only count for the
papers like i. As a highly cited paper with many citations, the out-degree of i can be ignored, compared with in-degree. So
we only consider the papers that cite i. Consider any two papers j, l belong to the zone of i and cite i as well. Further suppose
tj < tl − δ . If l is not a multidisciplinary paper, the probability of l citing j is pqβ(θj )e−α ltj −(1−α)ltl /(β(θi )e−α lti −(1−α)ltl ) ≈
pqe−α l(tj −ti ) , where we use the piecewise constant property of β(·) and make the assumption that the overlap of the zone j
(having a zone with probability q) is fully contained by that of i (this is justified if tj is large, which is common in the modeled
networks with T ≫ 1). If l is a multidisciplinary paper, the probability of l citing j should add max(0, E (f )− k+ (θl , tl ))/N (tl −
δ), where E (f ) is the expected value of f (k), and N (tl − δ) is the number of papers published until time tl − δ . The increment
is very small and could be ignored, so the probability of l citing j is approximately equal to pqe−α l(tj −ti ) .
Summing over all possible values of tj with the weight: the proportion of papers published in the tj th time unit and citing
i to all papers citing i, we obtain the approximate value of the expected LCC of i as follows
 
β(θi )mp
pqeα l(ti −tj )
T
ti +δ 2π
e−α lti −(1−α)ltj eltj dtj β(θi )(T − ti − δ)mp2 q
C (θi , ti ) ≈ = , (7)
k− (θ i , ti ) 2π k− (θi , ti )
 Z. Xie et al. / Physica A 456 (2016) 167–175 173

(a) DBLP 2013-09-29. (b) Modeled network.

Fig. 4. Local clustering coefficient as a function of in-degree, compared with the theoretical prediction of Eq. (7). The panels show the average local
clustering coefficient of k-degree nodes of two networks in Table 1 respectively.

Fig. 5. The tunable global clustering coefficients of modeled networks. Increasing the values of parameters p and q can increase the value of GCC. When
varying p, we set q = 0.5, and varying q, p = 1. The other parameters are listed in Table 1.

which is inversely proportional to the in-degree k− (θi , ti ). Eq. (7) also shows the parameters p and q can tune the value of
LCC, and so that of GCC (Fig. 5).

5. In- and out-assortativity

The empirical data show the highly cited papers tend to cite other papers also highly cited (Fig. 6(a)), which means the
citation networks are in-assortative. Tracking hot research topics is a reason for the in-assortativity. If a paper is highly
cited, it means that the topic discussed in this paper is a hot topic. Many researchers would focus on the topic, publish a lot
of papers about the topic, create a lot of citations between those papers, and then make those papers to be highly cited. The
empirical data are also out-assortative (Table 1, Fig. 6(c)). In fact, the annual average reference length of a paper increases
yearly, and papers usually tend to cite some papers newly published.
Simulations show modeled networks by our model are also in- and out-assortative (Table 1). To show how the model
does (Fig. 6(b), (d)), we derive the underlying formulae of the scaling relations between the published time of a paper i and
the average in- and out-degree of the papers citing and cited by i, N −,− (θi , ti ) and N +,+ (θi , ti ), respectively.
Firstly, we consider N −,− (θi , ti ). If i is a highly cited paper, the average in-degree of the papers citing i is mainly
contributed by some highly cited papers and can be calculated as follows
 
T β(θi )mp
t +δ
qk− (θi , s) 2π
e−α lti −(1−α)ls els ds β(θi )mpq
i
N −,−
(θi , ti ) ≈ ≈ (T − ti − δ), (8)
k− (θi , ti ) 2π
which is a decreasing function of ti so an increasing function of the in-degree of i. If i is not a highly cited paper (so has a
small in-degree) the papers citing i come from multidisciplinary papers. The probability of a paper simultaneously to be a
multidisciplinary and highly cited paper is small (which is qr) so the probability of N −,− (θi , ti ) having a large value is also
small. To sum up, the modeled networks are in-assortative. Secondly, we consider N +,+ (θi , ti ). If i is a non-multidisciplinary
174 Z. Xie et al. / Physica A 456 (2016) 167–175

(a) DBLP 2013-09-29. (b) Modeled network.

(c) DBLP 2013-09-29. (d) Modeled network.

Fig. 6. The means of the average in- and out-degree of the neighbors citing and cited by k in- and out-degree papers, N −,− (k) and N +,+ (k), respectively.
Panels (a,b) show those for DBLP 2013-09-29, and Panels (c,d) for the modeled network in Table 1.

paper, the average out-degree of the papers cited by i is

 
 ti −δ β(θi )mpq
1
k+ (θi , s) 2π
e−α ls−(1−α)lti els ds β(θi )mpq

e−(1−α)l ti − 1 − δ

N +,+ (θi , ti ) ≈ ≈ − , (9)
k+ (θi , ti ) 2π (1 − α)l e(1−α)l(ti −δ−1) − 1
which is an increasing function of ti so an increasing function of the out-degree of i. If i is a multidisciplinary paper,
N +,+ (θi , ti ) does not satisfy the result given by Eq. (9), because it cites some papers randomly. Simulations also show the
average out-degree of the papers cited by a paper with large out-degree fluctuates around a constant (Fig. 6(d)). In the model,
the large out-degree papers only count for multidisciplinary papers. Meanwhile, the number of multidisciplinary papers is
very small (which is achieved by giving a small value to r, e.g. r = 0.002) so the modeled networks are still out-assortative.

6. Conclusion

A directed geometric graph is proposed to model citation networks with exponentially growing nodes, which overcomes
certain shortcomings of our prior proposed model, and provides better reproductions of some typical statistical features of
the empirical data, e.g. out-degree distribution. The model potentially paves a geometric way to understand some aspects
of citation networks, e.g. it abstractly quantifies the academic influences of papers by geometric zones, expresses aging
of papers by decreasing the sizes of influences with the growth of time, and then interprets the power-law tails of in-
degree distributions of citation networks by the inhomogeneous influences of papers. Some shortcomings of the model are
indicative of the need for further research: how to geometrically realize that the in-degree distributions of papers published
in the same year are power-law; how to design a more realistic strategy of localizing the copying mechanism rather than
the randomly and uniformly selecting strategy used here. The last but most important problem that needs further research
is that, if to infer proper coordinates of a part of empirical data, how to infer the coordinates for the others and predict the
future, which needs the techniques in statistics and mapping knowledge domains.

Acknowledgments

We thank the editor and anonymous reviewer for their valuable suggestions and great help, and the hospitality of
Professor Alastair Spence in the centre for networks and collective behaviour in the University of Bath. This work is supported
 Z. Xie et al. / Physica A 456 (2016) 167–175 175

by the open fund from Key Laboratory of High Performance Computing (No. 2 01403-01), and the National University of
Defense Technology Graduate Teaching Reform Project (No. yjsy2013012).

References

[1] M.H. MacRoberts, B.R. MacRoberts, Problems of citation analysis, Scientometrics 36 (3) (1996) 435–444.
[2] D. Wang, C. Song, A.L. Barabási, Quantifying long-term scientific impact, Science 342 (2013) 127–131.
[3] T.A. Brooks, Evidence of complex citer motivations, J. Am. Soc. Inf. Sci. Technol. 37 (1) (1986) 34–36.
[4] S. Rednes, How popular is your paper? An empirical study of the citation distribution, Eur. Phys. J. B 4 (2) (1998) 131–134.
[5] F. Radicchi, S. Fortunato, A. Vespignani, Citation networks, in: A. Scharnhorst, K. Börner, P.V.D. Besselaar (Eds.), Models of Science Dynamics, Springer,
2012, pp. 233–257.
[6] F. Radicchi, C. Castellano, Understanding the scientific enterprise: citation analysis, data and modeling, in: Social Phenomena, Springer, 2015,
pp. 135–151.
[7] Price D.J. de Solla, Networks of scientific papers, Science 149 (3683) (1965) 510–515.
[8] Price D.J. de Solla, A general theory of bibliometric and other cumulative advantage process, J. Am. Soc. Inf. Sci. 27 (5) (1976) 292–306.
[9] D.J. Price, The scientific foundations of science policy, Nature 206 (1965) 233–238.
[10] F. Radicchi, S. Fortunato, B. Markines, A. Vespignani, Diffusion of scientific credits and the ranking of scientists, Phys. Rev. E 80 (5) (2009) 056103.
[11] L. Bornmann, H.D. Daniel, Universality of citation distributions-A validation of Radicchi et al.’s relative indicator cf = c /c0 at the micro level using
data from chemistry, J. Am. Soc. Inf. Sci. 60 (8) (2009) 1664–1670.
[12] T.S. Evans, N. Hopkins, B.S. Kaube, Universality of performance indicators based on citation and reference counts, Scientometrics 93 (2012) 473–495.
[13] J.R. Clough, J. Gollings, T.V. Loach, T.S. Evans, Transitive reduction of citation networks, J. Complex. Netw. 3 (2) (2015) 189–203.
[14] S.R. Goldberg, H. Anthony, T.S. Evans, Modelling citation networks, Scientometrics 105 (2015) 1577–1604.
[15] K.B. Hajra, P. Sen, Modeling aging characteristics in citation networks, Physica A 368 (2) (2005) 575–582.
[16] S.N. Dorogovtsev, J.F.F. Mendes, Evolution of networks with aging of sites, Phys. Rev. E 62 (2000) 1842–1845.
[17] Y.H. Eom, S. Fortunato, Characterizing and modeling citation dynamics, Plos One 6 (2011) e24926.
[18] P.L. Krapivsky, S. Redner, Network growth by copying, Phys. Rev. E 71 (2005) 036118.
[19] M.V. Simkin, P.V. Roychowdhury, Stochastic modeling of citation slips, Scientometrics 62 (2005) 367–384.
[20] Z.X. Wu, P. Holme, Modeling scientific-citation patterns and other triangle-rich acyclic networks, Phys. Rev. E 80 (3) (2009) 037101.
[21] G.J. Peterson, Steve Pressé S, K.A. Dill, Nonuniversal power law scaling in the probability distribution of scientific citations, Proc. Natl. Acad. Sci. USA
107 (2010) 16023–16027.
[22] M. Penrose, Random Geometric Graphs, in: Oxford Studies in Probability, 2003.
[23] M. Barthélemy, Spatial networks, Phys. Rep. 499 (2011) 1–101.
[24] Z. Xie, J. Zhu, D.X. Kong, J.P. Li, A random geometric graph built on a time-varying Riemannian manifold, Physica A 436 (2015) 492–498.
[25] Z. Xie, T. Rogers, Scale-invariant geometric random graphs, arXiv:1505.01332.
[26] Z. Xie, Z.Z. Ouyang, J.P. Li, A geometric graph model for coauthorship networks, J. Informetr. 10 (2016) 299–311.
[27] D. Krioukov, M. Kitsak, R.S. Sinkovits, D. Rideout, D. Meyer, M. Boguñá, Network cosmology, Sci. Rep. 2 (2012) 793.
[28] Z. Xie, Z.Z. Ouyang, P.Y. Zhang, D.Y. Yi, D.X. Kong, Modeling the citation network by network cosmology, Plos One 10 (2015) e0120687.
[29] J. Tang, J. Zhang, R.M. Jin, Z. Yang, K.K. Cai, L. Zhang, et al., Topic level expertise search over heterogeneous networks, Mach. Learn. J. 82 (2011) 211–237.
[30] KDD Cup 2003: Network mining and usage log analysis. http://www.cs.cornell.edu/projects/kddcup/datasets.html.
[31] P.C. Consul, G.C. Jain, A generalization of the Poisson distribution, Technometrics 15 (4) (1973) 791–799.
[32] A. Clauset, C.R. Shalizi, M.E.J. Newman, Power-law distributions in emprical data, SIAM Rev. 51 (2009) 661–703.