ArticlePDF Available

Predictive Effects of Structural Variation on Citation Counts

March 2012
Journal of the American Society for Information Science and Technology 63(3):431-449

Authors:

College of Computing and Informatics, Drexel University

A critical part of a scientific activity is to discern how a new idea is related to what we know and what may become possible. As the number of new scientific publications arrives at a rate that rapidly outpaces our capacity of reading, analyzing, and synthesizing scientific knowledge, we need to augment ourselves with information that can guide us through the rapidly growing intellectual space effectively. In this article, we address a fundamental issue concerning with what information may serve as early signs of potentially valuable ideas. In particular, we are interested in information that is routinely available and derivable upon the publication of a scientific paper without assuming the availability of additional information such as its usage and citations. We propose a theoretical and computational model that predicts the potential of a scientific publication in terms of the degree to which it alters the intellectual structure of the state of the art. The structural variation approach focuses on the novel boundary-spanning connections introduced by a new article to the intellectual space. We validate the role of boundaryspanning in predicting future citations using three metrics of structural variation, namely, modularity change rate, cluster linkage, and centrality divergence, along with more commonly studied predictors of citations such as the number of co-authors, the number of cited references, and the number of pages. Main

An overview of the structural variation model.

…

Scenarios that may increase or decrease individual terms in the modularity metric. [Color figure can be viewed in the online version, which is available at wiley.com.]

…

The depth and breadth of novel connections made by two articles with the highest modularity change rate. Both are review articles. Barilan 2008 (left) cited references in recent years within a relatively small number of clusters whereas Morris 2008 (right) cited many earlier publications across a wider range of clusters. [Color figure can be viewed in the online version, which is available at wiley.com.]

…

A close-up view of the lower right region shown in , showing novel links between distinct clusters of co-cited references. [Color figure can be viewed in the online version, which is available at wiley.com.]

…

Co-citations made by the five articles with the strongest Centrality Divergence scores in Terrorism (1996–2005) that reinforced existing patterns (left) and introduced novel connections (right). Structural variations were computed based on 2-year time slices. [Color figure can be viewed in the online version, which is available at wiley.com.]

…

Figures - uploaded by Chaomei Chen

Content may be subject to copyright.

Content uploaded by Chaomei Chen

Content may be subject to copyright.

Predictive Effects of Structural Variation on Citation

Counts

Chaomei Chen

College of Information Science and Technology, Drexel University, 3141 Chestnut Street, Philadelphia, PA 19104.

E-mail: chaomei.chen@drexel.edu

A critical part of a scientiﬁc activity is to discern how

a new idea is related to what we know and what may

become possible. As the number of new scientiﬁc pub-

lications arrives at a rate that rapidly outpaces our

capacity of reading, analyzing, and synthesizing scien-

tiﬁc knowledge, we need to augment ourselves with infor-

mation that can effectively guide us through the rapidly

growing intellectual space. In this article, we address a

fundamental issue concerning what kinds of informa-

tion may serve as early signs of potentially valuable

ideas. In particular, we are interested in information that

is routinely available and derivable upon the publication

of a scientiﬁc paper without assuming the availability

of additional information such as its usage and cita-

tions. We propose a theoretical and computational model

that predicts the potential of a scientiﬁc publication in

terms of the degree to which it alters the intellectual

structure of the state of the art. The structural varia-

tion approach focuses on the novel boundary-spanning

connections introduced by a new article to the intellec-

tual space. We validate the role of boundary-spanning in

predicting future citations using three metrics of struc-

tural variation—namely, modularity change rate, cluster

linkage, and Centrality Divergence—along with more

commonly studied predictors of citations such as the

number of coauthors, the number of cited references,

and the number of pages. Main effects of these factors

are estimated for ﬁve cases using zero-inﬂated negative

binomial regression models of citation counts. Key ﬁnd-

ings indicate that (a) structural variations measured by

cluster linkage are a better predictor of citation counts

than are the more commonly studied variables such

as the number of references cited, (b) the number of

coauthors and the number of references are both good

predictors of global citation counts to a lesser extent,

and (c) the Centrality Divergence metric is potentially

valuable for detecting boundary-spanning activities at

interdisciplinary levels.The structural variation approach

offers a new way to monitor and discern the potential

of newly published papers in context. The boundary-

spanning mechanism offers a conceptually simpliﬁed

Received September 1, 2011; revised October 3, 2011; accepted October 5,

2011

(wileyonlinelibrary.com). DOI: 10.1002/asi.21694

and unifying explanation of the roles played by commonly

studied extrinsic properties of a publication in the study

of citation behavior.

Introduction

A hallmark of scientiﬁc knowledge is its constant interplay

with new ideas proposed by the scientiﬁc community. New

ideas vary considerably in terms of how their value could be

perceived. Some are warmly embraced upon their conception

whereas others may go through lengthy periods of uncer-

tainty and controversy or even become totally ignored by

the scientiﬁc community. A critical part of scientiﬁc inquiry

is to discern where a new idea stands given what the sci-

entiﬁc community knows as a whole. This is a cognitively

demanding and conceptually challenging task. Not only do

we need to have an up-to-date understanding of the intel-

lectual structure of the relevant scientiﬁc ﬁelds but we also

must be able to identify exactly how a newly proposed idea

is connected to the intellectual structure. New scientiﬁc pub-

lications arrive faster than what any individual can possibly

read, analyze, and synthesize.

Detecting early signs of potentially valuable ideas has the-

oretical and practical implications. For instance, peer reviews

of new manuscripts and new grant proposals are under a grow-

ing pressure of accountability for safeguarding the integrity

of scientiﬁc knowledge and optimizing the allocation of

limited resources (Chubin, 1994; Chubin & Hackett, 1990;

Häyrynen, 2007; Hettich & Pazzani, 2006). Long-term strate-

gic science and technology policies require visionary think-

ing and evidence-based foresights into the future (Cuhls,

2001; Martin, 2010; Miles, 2010). In foresight exercises on

identifying future technology, experts’ opinions were found

to be overly optimistic with hindsight (Tichy, 2004). The

increasing specialization in today’s scientiﬁc community

makes it unrealistic to expect an expert to have a compre-

hensive body of knowledge concerning multiple key aspects

of a subject matter, especially in interdisciplinary research

areas.

Library (wileyonlinelibrary. com). DOI: 10.1002/asi.21694

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY, 63(3):431–449, 2012

432 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

The value, or perceived value, of an idea can be quanti-

ﬁed in many ways. For example, the value of a good idea can

be measured by the number of people’s lives that it has saved,

the number of jobs that it has created, or the amount of rev-

enue that it has generated. In the intellectual world, the value

of a good idea can be measured by the number of other ideas

that it has inspired or the amount of attention that it has drawn.

In this article, we are concerned with identifying patterns and

properties of information that can tell us something about the

potential values of ideas expressed and embodied in scientiﬁc

publications. A citation count of a scientiﬁc publication is the

number of times other scientiﬁc publications have referenced

the publication. Using citations to guide the search for rele-

vant scientiﬁc ideas by way of association, known as citation

indexing, was pioneered by Eugene Garﬁeld (1955). It is a

general consensus that citation behavior can be motivated

by both scientiﬁc and nonscientiﬁc reasons (Bornmann &

Daniel, 2006). Citation counts have been used as an indica-

tor of intellectual impact on subsequent research. There have

been debates over the nature of citations and whether all pos-

itive, negative, and self-citations should be treated equally.

Nevertheless, even a negative citation makes it clear that the

referenced work cannot be simply ignored.

What do we know about factors that may inﬂuence citation

counts in one way or another? One may address this ques-

tion from a few different points of view, largely depending

on where we draw our insights from: the past or the present.

An article that has been highly cited so far is likely to remain

highly cited according to the Matthew Effect (Merton, 1968).

An article that has been frequently downloaded or viewed

online is likely to become highly cited later (Brody & Harnad,

2005; Kurtz et al., 2005). Relying on direct evidence such as

visit counts, download counts, and citation counts that an

article already has obtained has relatively lower risks than

does making assessments based on indirect evidence. The

downside of such approaches is that the analysis is not pos-

sible until a sufﬁcient period of time elapses from the time

of publication so that the article has a reasonable exposure to

the scientiﬁc community.A much longer delay is expected to

conduct a citation analysis because of the longer life cycle of

scholarly publication.

Researchers have searched for other clues that may inform

us about the potential impact of a newly published scientiﬁc

paper, especially ones that can be readily extracted from rou-

tinely available information at the time of publication instead

of waiting for download and citation patterns to build up over

time. Factors such as the track record of authors, the pres-

tige of authors’ institutions, and the prestige of the journal in

which an article is published are among the most promis-

ing clues that can provide an assurance of the quality of

the article, to an extent (Boyack, Klavans, Ingwersen, &

Larsen, 2005; Hirsch, 2007; Kostoff, 2007; van Dalen &

Kenkens, 2005; Walters, 2006). The common assumption

central to approaches in this category is that great researchers

tend to continuously deliver great work, and along a simi-

lar vein, an article published in a high-impact journal also

is likely to be of high quality itself. On one hand, these

approaches avoid the reliance on data that may not be read-

ily available upon the publication of an article and thus free

analysts from constraints due to the lack of download and cita-

tion data. On the other hand, the sources of information used

in these approaches are indirect to the new ideas reported

in scientiﬁc publications. In an analogy, we give credit to

an individual based on his or her credit history instead of

directly assessing the risk of the current transaction. In such

approaches, we will not be able to know precisely where the

novelty of an idea is from or whether similar ideas have been

proposed in the past.

The approach that we will introduce in this article aims to

provide speciﬁc trails of evidence to show why and how an

idea is novel with reference to the current intellectual struc-

ture of a scientiﬁc domain. We conceptualize the development

of scientiﬁc knowledge as a process of interplay between the

intellectual structure and a stream of incoming new ideas

conveyed in newly published scientiﬁc papers. Each new

idea may alter the current intellectual structure or leave the

structure intact. The prediction of the potential value, or the

impact, of an idea can be made computationally in terms

of the degree of structural change introduced by the idea.

In this context, we call this approach a structural variation

model. For example, if a new idea connects previously dis-

parate patches of knowledge, then its transformative potential

is higher than is the potential of ideas that are limited to well-

trodden paths over the existing structure. The central idea

of this approach is a boundary-spanning mechanism, which

is conceptualized as a production rule that drives the cita-

tion process. The intellectual structure can be represented

by networks of ideas. The conceptual change brought by

newly published scientiﬁc articles to the intellectual structure

can be quantiﬁed based on information that comes with the

publication of such articles, notably the authorship and cited

references. To validate that the structural variation model does

capture insightful information about the potential value of an

article, we investigate the extent to which structural variation

measures of articles predict their subsequent citations along

with other more commonly studied predictors such as the

number of coauthors and the number of cited references. The

structural variation model offers a conceptually simple and

unifying explanation of several commonly identiﬁed citation

predictors. For example, review and survey articles are often

highly cited. An explanation in our model is that they tend

to synthesize individual areas in a broader context than do

original research articles and they are more likely to include

boundary-spanning connections, which could be both inten-

tional and unintentional when a large number of topics are

reviewed. The number of coauthors has been recognized as a

potential factor for predicting high citation counts. A possible

explanation could be that multiple coauthors bring different

areas of expertise, and as a result, boundary spanning takes

place as they collaborate.

In the following sections, we will describe the procedure

of constructing baseline representations of the intellectual

structure and deﬁne structural variation metrics. Then, we

will validate the role of structural variation mechanisms

433

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

in predicting subsequent citations through a series of gen-

eralized linear models, which are particularly suitable for

modeling count data with overdispersion and excessive zeros.

These concepts will be discussed later. We also will demon-

strate how the structural variation model provides a new

way to interact and explore speciﬁc connections that make

newly published scientiﬁc articles novel in the backdrop of

the current intellectual structure.

Structural Variation Model

There is a recurring theme from a diverse body of work on

creativity. A major form of creative work is to bridge previ-

ously disjoint bodies of knowledge. Notable studies include

the work of Ronald S. Burt (2004) in sociology and Don-

ald Swanson (1986a) in information science, and conceptual

blending as a theoretical framework for exploring human

information integration (Fauconnier & Turner, 1998). We

have been developing an explanatory and computational the-

ory of transformative discovery based on criteria derived from

structural and temporal properties (Chen, 2011; Chen et al.,

2009).

In the history of science, there are many examples of how

new theories revolutionized the contemporary knowledge

structure. For example, the 2005 Nobel Prize in medicine

was awarded to the discovery of Helicobacter pylori, a bac-

terium which was not believed to be possible to ﬁnd in a

human gastric system (Chen et al., 2009). In literature-based

discovery, Swanson (1986a) discovered previously unnoticed

linkage between ﬁsh oil and Reynaud’s syndrome. In terror-

ism research, before the September 11, 2001 terrorist attacks,

it was widely believed that only those who directly witness

a traumatic scene or directly experience a trauma could have

the risk of posttraumatic stress disorder (PTSD); however,

later research has shown that people may develop PTSD

syndromes by simply watching the coverage of a traumatic

scene on television (Chen, 2006). In drug discovery, one of the

major challenges is to ﬁnd new compound structures effective

in the vast chemical space that satisfy an array of constraints

(Lipinski & Hopkins, 2004). In mapping scientiﬁc frontiers

(Chen, 2003) and studies in science of science (Price, 1965),

it would be particularly valuable if scientists, funding agen-

cies, and policy makers could have tools that may assist them

to assess the novelty of ideas in terms of their conceptual

distance from the contemporary domain knowledge. In these

and many more scenarios, a common challenge for coping

with a constantly changing environment is to estimate the

extent to which the structure of a network should be updated

in response to newly available information.

Many studies have addressed factors that could explain

or even predict future citations of a scientiﬁc publication

(Aksnes, 2003; Hirsch, 2007; Levitt & Thelwall, 2008;

Persson, 2010). For example, is a paper’s citation count last

year a good predictor for new citations this year? Are the

download times a good predictor of citations? Is it true that the

more references a paper cites, the more citations it will receive

later? Similarly, the potential role of prestige, or the Matthew

Effect has been commonly investigated, ranging from the

prestige of authors to the prestige of journals in which articles

are published (Dewett & Denisi, 2004). However, many of

these factors are loosely and indirectly coupled with the con-

ceptual and semantic nature of the underlying subject matter

of concern. We refer to them as extrinsic factors in this article.

In contrast, intrinsic factors have direct and profound connec-

tions with the intellectual content and structure. One example

of an intrinsic factor is concerned with the structural variation

of a ﬁeld of study.A notable example is the work by Swanson

(1986a) on linking previously disjoint bodies of knowl-

edge, such as the connection between ﬁsh oil and Reynaud’s

syndrome.

Extrinsic Factors of Citations

Bornmann and Daniel (2006) reviewed 30 studies of citing

behavior published from the 1960s to mid-2005. The general

tendency they found in the empirical studies is that citing

behavior is motivated by many scientiﬁc and nonscientiﬁc

factors. Their review summarized the status of numerous

detailed questions studied in the past. Note that some of

the most highly cited articles are review articles or articles

about methodology or tools (Tijssen,Visser, & van Leeuwen,

2002). Several studies have conﬁrmed that the number of

references cited by a paper appears to be a good predictor

of its future citations, among other factors such as the pres-

tige of the paper’s author(s) and its journal (Walters, 2006).

For instance, review articles usually cite a large number of

references and tend to be highly cited (Aksnes, 2003). On

the other hand, the number of references cited by an article

alone is likely to be a poor proxy of the intellectual value of the

article. It is reasonable to argue that which references an arti-

cle cites matters more than how many references it cites. For

example, if an article Acites considerably more references

than does another Article B, then there are a few plausible

explanations:

•Article Aaddresses a topic much more extensively than does

Article B.

•Article A, for example, a review paper, addresses more topics

than does Article B.

•Article A, for example, a groundbreaking research paper,

addresses a topic that builds on a synthesis of multiple topics

whereas Article Bsynthesizes a less number of topics.

The list can go on. The value of our approach is to take

into account structural variations introduced by an article to

provide additional insights into the relationship between an

article and the state of the art.

Boyack et al. (2005) reported a method for predicting

the importance of current papers based on journal impor-

tance, reference importance, and author reputation. Measures

of importance in their study were based on citations to

780,049 papers published in 2002, with a window of citation

from the beginning of 2002 till the end of 2003. Forty-

eight percent of papers were never cited in this window.

The journal importance was calculated using the formula

published by the Institute for Scientiﬁc Information (ISI).

434 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

The reference importance was calculated as the number of

times the reference in question was cited by papers pub-

lished in 2002. The author reputation was calculated as the

frequency of appearances of author–journal pairs in a 4-year

window before 2002. According to their report, a regression

analysis of logarithmic-transformed variables accounted for

approximately 30% of variance.Among the three importance

variables, the journal impact had the strongest correlation.

They suggested that properties such as these importance

factors can be used to rank articles without the need to

wait for citations to accumulate. While the ranking proce-

dure that they suggested is useful, their procedure implies

an assumption that articles citing more popular references

may get more citations. In this article, we take a different

approach that focuses speciﬁcally on how the contemporary

intellectual structure may change as a result of the way a

new article links references. In other words, the argument

that articles get more citations because they cite popular

references only reveals part of the story. Is it possible that

frequent citations to an article are not because of the popu-

larity of references it cites but rather where these references

are located in the intellectual structure of knowledge? Fur-

thermore, is it possible to measure the value of a scientiﬁc

contribution in terms of how many new ways of thinking it

may introduce? This is indeed a major motivating question of

our study.

Walters (2006) studied citations received by 428 articles

published in 12 crime-psychology journals in a single year of

2003 and identiﬁed nine major predictors of citations, includ-

ing author characteristics such as gender and citations to ﬁrst

authors’ publications 2 years prior to the new publication;

article characteristics such as coauthors, article length, and

subject matter; and journal characteristics such as journal

impact. Walters’ study used negative binomial regression of

the citations received by these 428 articles 1 or 2 years after

their publication. His study suggested that author impact may

be a more powerful predictor of citations than is the impact

of a journal.

Kostoff (2007) compared highly and poorly cited research

articles published in The Lancet and found that the most cited

articles tend to have more coauthors, cite more references, and

have a longer abstract and more pages. Highly cited papers

tend to report clinical trials of much larger sample sizes than

do poorly cited papers. In a speciﬁc context, the h-index was

found to have a strong correlation with the number of cita-

tions: The h-index calculated based on the ﬁrst 12 years in

a sample dataset has a correlation coefﬁcient of 0.60 with

citations in the second 12 years (Hirsch, 2007).

Levitt and Thelwall (2008) studied highly cited articles in

six subject areas and found that predicting citation rank-

ing of highly cited articles using the subtotals of citations

in Years 5 and 6 is more accurate than is using the total

of citations in the ﬁrst 6 years. Skilton (2009) found that

frequently cited coauthors and authors with diverse dis-

ciplinary backgrounds tend to be highly cited in natural

sciences. In contrast, the variety of disciplines represented

collectively by the authors of an article does not inﬂuence

citations. Persson (2010) addressed the question of whether

highly cited papers are more international (i.e., written by

coauthors from different geographic locations) than are less

highly cited ones. Based on data from four research areas,

three universities, four cities, and two countries, he con-

cluded that international papers dominate highly cited papers

from small countries, but are not well represented overall in

high-impact papers.

In addition to characteristics of authors and their institu-

tions, features of articles such as the length of the title, the

number of ﬁgures and tables, and the number and recency of

references and research methodology have been studied in

the literature (Haslam et al., 2008).

Intrinsic Factors of Citations

Researchers have made various attempts to characterize

future citations and identify emerging core articles (Shibata,

Kajikawa, & Matsushima, 2007; Walters, 2006). Shibata et

al. (2007) for example, studied citation networks in two

subject areas, Gallium Nitride and Complex Networks, and

found that while past citations are a good predictor of near-

future citations, the betweenness centrality is correlated with

citations in a longer term.

Upham, Rosenkopf, and Ungar (2010) studied the role of

cohesive intellectual communities—schools of thought—in

promoting and constraining knowledge creation. They ana-

lyzed publications on management, and concluded that it is

signiﬁcantly beneﬁcial for new knowledge to be a part of a

school of thought and that the most inﬂuential position within

a school of thought is in the semiperiphery of the school.

In particular, boundary-spanning research positioned at the

semiperiphery of a school would attract attention from other

schools of thought and receive the most citations overall.

Their study used a zero-inﬂated negative binomial regres-

sion (ZINB). Negative binomial regression models have been

used to predict the expected mean patent citations (Fleming &

Bromiley, 2000). Hsieh (2011) studied inventions as a com-

bination of technological features. In particular, the closeness

of features plays an interesting role. Neither overly related nor

loosely related features are good candidates for new inven-

tions. Useful inventions arise with rightly positioned features

where the cost of synthesis is minimized.

Takeda and Kajikawa (2010) reported three stages of clus-

tering in citation networks. In the ﬁrst stage, core clusters

are formed, followed by the formation of peripheral clus-

ters and the continuous growth of the core clusters. Finally,

the core clusters’ growth again becomes predominant. Buter,

Noyons, and van Raan (2011) studied the emergence of an

interdisciplinary research area from ﬁelds that had not shown

interdisciplinary connections before. They used journal sub-

ject categories as a proxy for ﬁelds and citations as a measure

of interdisciplinary connection.

Lahiri, Maiya, Sulo, Habiba, and Wolf (2008) addressed

how structural changes of a network may inﬂuence the spread

of information over the network.Although they did not study

bibliographic networks, per se, their study has indicated that

435

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

FIG. 1. An overview of the structural variation model.

predictions made about how information spreads over a net-

work are sensitive to structural changes of the network. This

observation underlines the importance of taking structural

change into account in the development of metrics based on

topological properties of networks.

Leydesdorff (2001, p. 146) raised questions that are

closely related to what we are addressing: “How does the

new text link up to the literature, and what is its impact

on the network of previously existing relations?” He took

a quite different approach and analyzed word occurrences

in scientiﬁc papers from an information-theoretic perspec-

tive. In his approach, the publication of a paper is perceived

as an event that may lead to the reduction of uncertainty

involved in the current state of knowledge. He devised

diagrams that depict pathways of how a particular paper

improves the efﬁciency of communication. Although the

information-theoretic approach and our structural variation

approach currently operate on different units of analysis

with distinct theoretical underpinnings, both share the fun-

damental concern of changes introduced by newly published

scientiﬁc papers on the existing body of knowledge.

As noted, many studies in the literature have addressed

factors that may inﬂuence citations. The value of our work

is the introduction of the structural variation paradigm along

with computational metrics that can be integrated into inter-

active exploration systems to better understand precisely the

impact of individual links made by a new article.

In this article, we conceptualize that structural variation

is an essential process that advances scientiﬁc knowledge.

The intellectual structure at a given point of time is subject

to structural changes introduced by newly published scien-

tiﬁc articles. The intellectual structure may be represented

in many different types of networks, including networks of

co-cited references, networks of co-cited authors, or net-

works of co-occurring keywords. Given a scientiﬁc article,

publications prior to the publication of the article form a

baseline representation of the intellectual structure at the spe-

ciﬁc time of publication. Structural variation metrics measure

the degree of structural change introduced by information

derived from the article. An overview of the procedure is

depicted in Figure 1. Individual steps are explained later in

corresponding sections.

436 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

Data Collection

First, we choose a topic or a target ﬁeld of study. In this

article, we include ﬁve cases drawn from four research areas:

terrorism, mass extinction, complex network analysis, and

knowledge domain visualization. We have studied some of

these research areas in our previous studies. Choosing a famil-

iar domain can help us to concentrate on new insights to be

revealed by our new approach. We also include a dataset of

articles deﬁned by the journals in which they are published.

We used the Web of Science as the source of our data, although

the method is applicable to other sources of citation data, such

as Scopus and Google Scholar. In the rest of this article, bib-

liographic records were retrieved from the Web of Science

unless stated otherwise.

Given a chosen topic, a dataset can be constructed in two

ways. One way is to search for records of articles by rele-

vance, determined by matching terms and other attributes.

For instance, a dataset of bibliographic records on terrorism

research can be constructed by using a topic search for the

occurrence of the term terrorism in the title, the abstract,

or the keyword ﬁeld of a record. It is possible to deﬁne a

dataset based on speciﬁc source journals, such as all articles

published in Nature and Science. The other way is to locate

articles by association formed by cited references. The sec-

ond method starts with a seed article. All articles that have

at least one overlapping reference with the seed article are

identiﬁed as related records. The global citation count of an

article is the total number of times the article is referenced

by other articles in the Web of Science at the time of retrieval.

The local citation count is the total number of times the arti-

cle is referenced by other articles within the retrieved dataset,

which is a subset of the entire collection of the Web of Science.

The global citation count is used in the analysis reported in

this article because we assume that it is a less biased proxy

of impact than is the local citation count.

A seed article is used as a starting point to reconstruct the

structural representations prior to and after the publication of

the seed article. We use a two-stage expansion process. First,

all the references cited by the seed articles are identiﬁed in

the Web of Science as a set CitedBy(s). Second, articles in the

Web of Science that cite at least one reference in CitedBy(s)

form a set SExpanded(s).

SExpanded(s ∈Articles)

={a∈Articles|CitedBy(a) ∩CitedBy(s) =∅}

The ﬁnal set of articles serves as the representation of the

population of the original seed article. This process of

citation expansion can be repeated as many times as mean-

ingful to the purpose of the intended study; for example,

∪a∈SExpanded(s) SExpanded (a). A description of the process is

detailed in our earlier publication on forward expansion

(Chen, Lin, & Zhu, 2006). Although we do not rule out the

possibility that a relevant article does not have any overlap

with an expanded set of articles generated in this procedure

(e.g., the relevance is purely established at a semantic level),

in practice it would be rare to come across an article that is

relevant, but cites a totally different body of literature, espe-

cially when the body of literature is identiﬁed by the two-step

citation expansion. Such articles would either ignore the rele-

vant literature or completely innovative. Two cases are based

on seeded datasets—Complex Network Analysis and CiteS-

pace Expanded—whereas the other three cases are based on

the relevance of topic search.

To explore various conﬁgurations of the structural varia-

tion model, we apply the procedure to a number of datasets

representing a diverse range of topics and subject domains.

The background of each case is outlined, and additional

references are provided for further reading.

•Complex Network Analysis

This dataset overlaps with the Small-World Networks dataset,

but the two datasets were constructed differently. A topic

search for “complex network analysis” revealed two most

frequently cited articles: Barabási and Albert (1999) and

Watts and Strogatz (1998).We constructed an updated dataset

with these two articles as seed articles. First, one seed arti-

cle was used to form a subdataset. Then, the two datasets

were merged to form a combined dataset on complex network

analysis.

The seed article (Barabási & Albert, 1999) has the high-

est citation count. As of August 5, 2011, it has been cited

5,792 times in the Web of Science. The article by Watts

and Strogatz (1998) has the second-highest citation count

of 5,291. For each of the seed articles, we retrieved all

the records that share at least one cited reference with the

seed article. Then, we merged all the records and formed

a combined dataset. The seed Barabási_1999 has 8,919

related articles published between 1980 and 2011. If we limit

our search to the period of 1996 to 2004 and to original

research articles, reviews, and proceedings, 2,326 records

were retained in the subset. The seed article Watts_1998 has

14,393 related records. Interestingly, this is 1.6 times more

than what Barabási_1999 has, although Barabási_1999 has

500 more citations. It appears that Watts_1998 involves a

wider range of topics than does the Barabási_1999 article. The

merged dataset from the two subsets contains 6,764 records.

We ﬁrst analyzed the dataset seeded by Barabási1999, then

the dataset seeded by Watts1998, and ﬁnally, the merged

dataset.

•Mass Extinctions research is not new. Some of its major

groundbreaking works appeared in early 1980s. On the other

hand, the ﬁeld is as active as ever because many questions

remain concerning mass extinctions. Our previous study of

the evolution of mass extinctions research between 1981 and

2004 revealed a shift of research focus in early 1990s from

the K-T boundary mass extinction to the Permian extinction

(Chen, 2006; Chen, Cribbin, Macredie, & Morar, 2002). We

expect to detect considerable structural variations within the

range of 1991 to 2010.

The Mass Extinctions dataset contains 1,745 records of

journal articles, conference proceedings, and review arti-

cles published between 1991 and 2010 on mass extinctions.

Records were retrieved using a topic search query “mass

extinction” on April 2010. The dataset does not cover articles

published later in 2010, but this does not affect our analysis

because we take the exposure time into account in generalized

linear models.

437

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

•Terrorism

This dataset consists of bibliographic records that resulted

from a topic search in the Web of Science on terrorism between

1996 and 2005. A total of 1,303 articles were retrieved on

July 11, 2011. The citation counts of articles in the dataset

indicate the number of citations an article has received since

its publication up to July 2011. For example, an article pub-

lished in January 1996 would have the longest exposure time

of over 15 years whereas an article published in December

2005 would have the shortest exposure time of more than 5

years. The exposure time is taken into account in the construc-

tion of generalized linear models (discussed in more detail

later).

In a previous study, we analyzed an earlier period of the

same subject (1990–2003) (Chen, 2006). Several conceptual

transformations were revealed by our previous study. In par-

ticular, we were able to identify a conceptual change from

the topological properties of a synthesized network of co-

cited references. Prior to the terrorist attacks on September

11, 2001, it was generally believed that PTSD would be

found only in people who had a direct experience of trauma.

Researchers then discovered after the extensive coverage of

the September 11 terrorist attacks on mass media that peo-

ple who were far away from New York also could develop

PTSD symptoms (Galea et al., 2002). This is one example of

boundary-spanning in that a new idea redeﬁnes the boundary

of a topic area and introduces fundamental changes to the orig-

inal intellectual structure. The terrorist attacks on September

11, 2001 occurred within the 1996 to 2005 range. Therefore,

we expect that the updated dataset would provide relevant

information for us to study the dynamics of the structural

variation.

Two cases are derived from this dataset with different con-

ﬁgurations of the length of time slices. We use both 2- and

3-year time slices to identify the procedural implications of

using time slices of different duration.

•CiteSpace Expanded

The case is seeded by our 2006 article on CiteSpace (Chen,

2006). This dataset consists of papers that have at least one

cited reference in common with our seed article. Based on

the content of the seed article, we anticipate that there will be

three subareas involved in the dataset: scientometrics, mass

extinctions, and terrorism. CiteSpace is a scientometric tool

that has been applied to the analysis of mass extinctions and

terrorism.

Baseline Networks

The basic assumption in the structural variation approach

is that the extent of a departure from the current intellectual

structure is a necessary condition for a potentially transforma-

tive idea in science. In other words, a potentially transforma-

tive idea ﬁrst needs to bring changes to the existing structure

of knowledge. To measure the degree of structural variation

introduced by a scientiﬁc article, the intellectual structure at

a particular moment of time needs to be represented in such a

way that structural changes can be computationally detected

and manually veriﬁable. Bibliographic networks can be com-

putationally derived from scientiﬁc publications. Research

in scientometrics and citation analysis routinely uses cita-

tion and co-citation networks as a proxy of the underlying

intellectual structure. In this article, we focus on using sev-

eral types of co-citation and co-occurrence networks as the

representation of a baseline network.

A network represents how a set of entities are connected.

Entities are represented as nodes, or vertices, in the network.

Their connections are represented as links, or edges. Relevant

entities in our context include several types of information

that can be computationally extracted from a scientiﬁc arti-

cle, such as references cited by the article, authors and their

afﬁliations, the journal in which the article is published, and

keywords in the article. In this article, we limit our discus-

sions to networks that are formed with a single type of entity,

although networks of multiple types of entities are worth con-

sidering once we establish a basic understanding of structural

variations in networks of a single type of entity.

Once the type of entities is chosen, the nature of the inter-

connectivity between entities is to be speciﬁed to form a

network. Networks of co-occurring entities represent a wide

variety of types of connectivity. A network of co-occurring

words represents how words are related in terms of whether

and how often they appear in the vicinity of each other. Co-

citation networks of entities such as references, authors, and

journals can be seen as a special case of co-occurring net-

works. For example, co-citation networks of references are

networks of references that appear together in the bodies of

scientiﬁc papers; these references are co-cited.

Networks of co-cited references represent more speciﬁc

information than do networks of co-cited authors because

references of different articles by the same author would be

lumped together in a network of co-cited authors. Similarly,

networks of co-cited references are more speciﬁc than are

networks of co-cited journals. We refer to such differences

in speciﬁcity as the granularity of networks. Measurements

of structural variation need to take the granularity factor into

account because it is reasonable to expect that networks at dif-

ferent levels of granularity would lead to different measures

of structural variations.

Another decision to be made about a baseline network is

a sampling issue. Taking a particular year as a standing point

to look at in the past, how far back should we consider in

the construction of a baseline network that would adequately

represent the underlying intellectual structure? Does the net-

work become more accurate if we go back more into the past?

Will it be more efﬁcient if we limit it to the most recent years

that really matter the most? In this article, given articles pub-

lished in a particular year Y, the baseline network represents

the intellectual structure using information from articles pub-

lished up to year Y−1. Two types of baseline networks are

investigated in this article: ones using a moving window of

a ﬁxed size [Y−k,Y−1] and ones using the entire history

[Y0,Y−1], where Y0is the earliest year of publication for

records in the given dataset.

Structural Variation Metrics

We expect that the degree of structural variation intro-

duced by a new article can offer prospective information

438 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

because of the boundary-spanning mechanism. If an article

introduces novel links that span the boundaries of different

topics, then we expect that this signiﬁes its potential in taking

the intellectual structure for a new turn.

Given a baseline network, structural variationscan be mea-

sured based on information provided by a particular article. In

this article, we introduce three metrics of structural variation.

Each metric quantiﬁes the degree of change in the baseline

network introduced by information provided by an article.

No usage data are involved in the measurement. The three

metrics are modularity change rate, intercluster linkage, and

Centrality Divergence.The deﬁnitions of the ﬁrst two metrics

depend on a partition of the baseline network, but the third one

does not. A partition of a network decomposes the network

into nonoverlapping groups of nodes. For example, clustering

algorithms such as spectral clustering can be used to partition

a network.

Principles

The theoretical underpinning of the structural variation

is that scientiﬁc discoveries, at least a subset of them, can

be explained in terms of boundary-spanning, brokerage, and

synthesis mechanisms in an intellectual space (Chen et al.,

2009). This conceptualization generalizes the principle of

literature-based discovery pioneered by Swanson (1986a,

1986b), which assumes that connections between previously

disparate bodies of knowledge are potentially valuable. In

Swanson’s famous ABC model, the relationships AB and

BC are known in the literature. The potential relationship

AC becomes a candidate that is subject to further scientiﬁc

investigation (Weeber, 2003). Our conceptualization is more

generic in several ways. First, in theABC model, theAC rela-

tion changes an indirect connection to a direct connection

whereas our structural variation model makes no assump-

tion about any prior relations. Second, in the ABC model,

the scope of consideration is limited to relationships involv-

ing three entities. In contrast, our structural variation model

takes a wider context into consideration and addresses the

novelty of a connection that links groups of entities as well as

connections linking individual entities. Because of the broad-

ened scope of consideration, it becomes possible to more

effectively search for candidate connections. In other words,

given a set of entities, the size of the search space of poten-

tial connections can be substantially reduced if additional

constraints are applicable for the selection of candidate con-

nections. For example, the structural hole theory developed

in social network analysis emphasizes the special potential of

nodes that are strategically positioned to form brokerage, or

boundary-spanning, links and create good ideas (Burt, 2004;

Chen et al., 2009).

Modularity Change Rate

Given a partition of a network (i.e., a conﬁguration of

clusters), the modularity of the network measures the degree

of interconnectivity among the groups of nodes identiﬁed

by the partition. If different clusters are loosely connected,

then the overall modularity would be high. In contrast, if

clusters are interwoven, then the modularity would be low. We

follow Newman’s algorithm (Newman, 2006) to calculate the

modularity with reference to a cluster conﬁguration generated

by spectral clustering (Chen, Ibekwe-SanJuan, & Hou, 2010;

Luxburg, 2006). Suppose the network Gis partitioned by a

partition Cinto kclusters such that G=c1+c2+…+ck,

Q(G) is deﬁned as follows, where mis the total number of

edges in the network G, and nis the number of nodes in G.

δ(ci,c

j) is known as Kronecker’s delta. It is 1 if nodes niand

njbelong to the same cluster and 0 otherwise. deg(ni)isthe

degree of node ni. The range of Q(G) is between −1 and 1.

Q(G, C) =1



i,j =0[δ(ci,c

j)·(A]ij −deg(ni)·deg(nj)

2m

The modularity of a network is a measure of the overall struc-

ture of the network. Its range is between −1 and 1. The

modularity change rate (MCR) of a scientiﬁc paper measures

the relative structural change due to the information from the

published paper with reference to a baseline network. For

each article a, and a baseline network Gbaseline, we deﬁne the

MCR as follows:

MCR(a) =Q(Gbaseline,C)−Q(Gbaseline⊕Ga,C)

Q(Gbaseline,C) ·100

where Gbaseline ⊕Gais the updated baseline network by

information from the article a. For example, suppose refer-

ence nodes niand njare not connected in a baseline network

of co-cited references but they are co-cited by article a. A new

link between niand njwill be added to the baseline network.

In this way, the article changes the structure of the baseline

network.

Intuitively, adding a new link anywhere in a network

should not increase the modularity of the network; it should

either reduce the modularity or leave it intact. However, the

change of modularity is not a monotonic function as we ini-

tially expected. In fact, it depends on where the new link is

added and how the network is structured. Adding a link may

reduce the proportion of the modularity in some clusters, but

it may increase the modularity in other clusters in the net-

work. Thus, the overall modularity change is not monotonic.

Diagrams in Figure 2 illustrate these scenarios.

Without losing any generality, assume that an article adds

one link at a time to a given baseline network. If the new

link connects two distinct clusters, then it has no effect on

the corresponding term in the updated modularity because

by deﬁnition, δij =0, and the corresponding term becomes 0.

Such a ink is illustrated by the dashed link e5,10 in the top dia-

gram in Figure 2. The new link, eij , will increase the degree

of nodes iand jby 1 (i.e., deg(i) will become deg(i) +1.

The total number of edges mwill increase to m+1. A simple

calculation at the bottom of Figure 2 shows that terms in the

modularity formula involving blue links will decrease from

their previous values. However, if the network has clusters

439

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

FIG. 2. Scenarios that may increase or decrease individual terms in the modularity metric. [Color ﬁgure can be viewed in the online version, which is

available at wiley.com.]

such as CAwith no changes in node degrees, then the corre-

sponding values of terms of lines in red will increase from

their previous values as the denominator increases from 2mto

2(m+1). In summary, the updated modularity may increase

as well as decrease, depending on the structure of the net-

work and where the new link is added. With this particular

deﬁnition of modularity, between-cluster links are always

associated with a zero-valued term in the overall modular-

ity formula due to the Kronecker’s delta. What we see in the

change of modularity is a combination of results from several

scenarios that are indirectly affected by the newly added link.

Next, we will introduce our next metric to reﬂect the changes

in terms of between-cluster links.

As shown in Figure 3, two articles with the highest MCRs

in a small dataset of papers that cited Chen (2006) were

review articles. Review articles tend to cover many top-

ics, as one would expect, which would mean adding many

boundary-spanning co-citation connections to baseline net-

works and thus leading to a substantial structural variation

in terms of modularity. The two review papers have dif-

ferent distributions of boundary spanning links: one with

references concentrated on a small number of clusters in

recent years and the other with references traced back sev-

eral decades ago in distinct clusters. Both papers cited the

informetrics cluster (No. 11).

Cluster Linkage

The Cluster Linkage (CL) measures the overall structural

change introduced by an article ain terms of new connections

added between clusters. Its deﬁnition assumes a partition of

the network. We introduce a function of edges λ(ci,c

j) which

is the opposite of δij used in the modularity deﬁnition. The

value of λij is 1 for an edge across distinct clusters ciand

cj. It will be 0 for edges within a cluster. λij will allow us

to concentrate on between-cluster links and ignore within-

cluster links, which is the opposite of how the modularity

metric is deﬁned. The new metric Linkage is the sum of all

440 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

FIG. 3. The depth and breadth of novel connections made by two articles with the highest modularity change rate. Both are review articles. Barilan 2008

(left) cited references in recent years within a relatively small number of clusters whereas Morris 2008 (right) cited many earlier publications across a wider

range of clusters. [Color ﬁgure can be viewed in the online version, which is available at wiley.com.]

the weights of between-cluster links eij divided by K—the

total number of clusters in the network. Linking to itself is

not allowed (i.e., we assume eii =0 for all nodes. Using link

weights makes the metric sensitive to links that strengthen

connections between clusters in addition to novel links that

make unprecedented connections between clusters.

It is possible to take into account the size of clus-

ters that a link is connecting so that connections between

larger sized clusters become more prominent in the mea-

surement. For example, one option is to multiple each eij

by √(size(c↓i) ·size(c↓j))/ max(size(c↓ki)). In this article,

the metric is deﬁned without such modiﬁcations for simplic-

ity. Suppose that Cis a partition of G; then the Linkage metric

is deﬁned as follows:

Linkage(G, C) =n

i=jλij eij

λij =0,n

i∈cj

1,n

i∈ cj

The Cluster Linkage is deﬁned as the difference of Link-

age before and after new between-clusters links added by an

article a.

CL(a) =Linkage(a)

=Linkage(Gbaseline ⊕Ga,C)−Linkage(G baseline,C)

Linkage(G+G) is always greater than or equal to Link-

age(G). Thus, CL is nonnegative.

Centrality Divergence

The Centrality Divergence (CKL) metric measures the

structural variation caused by an article ain terms of the

divergence of the distribution of betweenness centrality

CB(vi) of nodes viin the baseline network. This deﬁni-

tion does not involve any partitions of the network. If

nis the total number of nodes, the degree of structural

change CKL(G,a) can be deﬁned in terms of the KL

divergence.

CKL(Gbaseline ,a) =



i=0

pi·logpi

qi

pi=CB(vi,G

baseline)

qi=CB(vi,G

updated)

For nodes where pi=0orqi=0, we reset them as a small

number 10−6to avoid log(0).

Statistical models. We constructed negative binomial (NB)

and ZINB models to validate the role of structural variation

in predicting future citation counts of scientiﬁc publica-

tions. The negative binomial distribution is generated by a

sequence of independent Bernoulli trials. Each trial is either

a “success” with a probability of por a “failure” with a

probability of (1 −p). Here, the terminology of success and

failure in this context does not necessarily represent any prac-

tical preferences. The random number of successes Xbefore

441

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

encountering a predeﬁned number of failures rhas a negative

binomial distribution:

X∼NB(r, p).

One can adapt this deﬁnition to describe a wide variety of

count events. Citation counts belong to a type of count events

with an overdispersion (i.e., the variance >the mean). NB

models are commonly used in the literature to study this type

of count events. Two types of dispersion parameters are used

in the literature, θand α, where θ·α=1.

Zero-inﬂated count models are commonly used to account

for excessive zero counts (Hilbe, 2011; Lambert, 1992). Zero-

inﬂated models include two sources of zero citations: the

point mass at zero I{0}(y) and the count component with a

count distribution fcount(counts) such as negative binomial

or Poisson (Zeileis, Kleiber, & Jackman, 2011). The proba-

bility of observing a zero count is inﬂated with probability

π=fzero(zero citations).

fzero-inﬂated(citations)=π×I{0}(citations)+(1−π)

×fcount(citations).

ZINB models are increasingly used in the literature to

model excessive occurrences of zero citations (Fleming &

Bromiley, 2000; Upham et al., 2010). The report of a ZINB

model consists of two parts: the count model and the zero-

inﬂated model. One way to test whether a ZINB model is

superior to a corresponding NB model is known as the Vuong

test. The Vuong test is designed to test the null hypothesis that

the two models are indistinguishable. Akaike’s Information

Criterion (AIC) also is commonly used to evaluate the good-

ness of a model. Models with lower AIC scores are regarded

as better models in terms of the relative goodness of ﬁt.

In this article, we focus on global citation counts of scien-

tiﬁc publications recorded in the Web of Science. NB models

are deﬁned as follows using log as the link function.

Global citations ∼Coauthors +Modularity Change Rate

+Cluster Linkage +Centrality Divergence

+References +Pages.

Global citations is the dependent variable. Coauthors is

a factor of three levels of 1, 2, and 3. Level 3 is assigned to

articles with three or more coauthors. Coauthors is an indi-

rect indicator of the extent to which an article synthesizes

ideas from different areas of expertise represented by each

coauthor.

Three structural variation metrics are included as covari-

ants in generalized linear models: MCR, CL, and CKL.

According to our theory of creativity, groundbreaking ideas

are expected to cause strong structural variations. If global

citation counts provide a reasonable proxy of recognitions

of intellectual contributions in a scientiﬁc community, we

would expect that at least some of the structural variation met-

rics will have statistically signiﬁcant main effects on global

citations.

The number of cited references and the number of pages

are commonly reported in the literature as good predictors

of citations. To compare the effects of structural variation

with these commonly reported extrinsic properties of sci-

entiﬁc publications, References and Pages are included in

the models. Our theory offers a simpler explanation why the

more references a paper cites, the more citations it appears

to get. Due to the boundary-spanning synthetic mechanism,

an article needs to explain multiple parts and how they can

be innovatively connected. This process will result in citing

more references than would an article that covers a narrower

range of topics. Review papers by their nature belong to this

category.

It is known that articles published earlier tend to have

more citations than do articles published later. The exposure

time of an article is included in the NB models in terms of a

logarithmically transformed year of publication of an article.

An intuitive way to interpret coefﬁcients in NB models is

to use incidence rate ratios (IRRs) estimated by the models.

For example, if Coauthors has an IRR of 1.5, it means that

as the number of coauthors increases by 1, the global citation

counts would be expected to increase by a factor of 1.5 (i.e.,

increasing 1.5 times) while holding other variables in the

model constant. In our models, we will particularly examine

statistically signiﬁcant IRRs of structural variation models.

ZINB models use the same set of variables. The count

model of ZINB is identical to the NB model described ear-

lier. The zero-inﬂated model of the ZINB uses the same set

of variables to predict the excessive zeros. We found little

in the literature about good predictors of zeros in a com-

parable context. We choose to include all six variables in

the zero-inﬂated model to provide a broader view of the

zero-generating process. ZINBs are deﬁned as follows:

Global citations ∼Coauthors +Modularity Change Rate

+Cluster Linkage +Centrality Divergence

+References +Pages.

Zero citations ∼Coauthors +Modularity Change Rate

+Cluster Linkage +Centrality Divergence

+References +Pages.

Results

Unless stated otherwise, baseline networks are formed

with a 2-year sliding window [Y −2, Y −1] for papers pub-

lished in year Y. Co-citation links made by the top-100 most

cited articles in each year of the sliding window are used to

construct a baseline network.

•Complex Network Analysis (1996–2004)

Figure 4 shows an overview of a network of co-cited ref-

erences in this ﬁeld. Only major clusters’ labels are shown

in the ﬁgure. The largest cluster (Cluster 8) is labeled com-

plex network. The groups of clusters in the lower left of

the visualization include social capital (Cluster 12) and old-

helping-partner (Cluster 11). The groups of clusters in the

lower right of the visualization include striate cortical activ-

ity (Cluster 2), stochastic resonance (Cluster 1), and phase

442 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

FIG. 4. A network of co-cited references derived from the Complex NetworkAnalysis (1996–2004). Dashed lines in red are novel connections made by the

groundbreaking article (Watts & Strogatz, 1998), which has the highest scores in both Cluster Linkage (5.43) and Centrality Divergence (1.14). [Color ﬁgure

can be viewed in the online version, which is available at wiley.com.]

synchronization (Cluster 5). Dashed lines in red are novel con-

nections made by Watts and Strogatz (1998) at the time of its

publication. The article has the highest scores in CL and CKL

scores; 5.43 and 1.14, respectively. The ﬁgure offers a visual

conﬁrmation that the article was indeed making boundary-

spanning connections. Recall that the dataset was constructed

by expanding the seed article based on forward citation links.

These boundary-spanning links provide empirical evidence

that the groundbreaking paper was connecting two groups of

clusters. The emergence of Cluster 8 complex network was

the consequence of the impact.

Figure 5 is a close-up view of the lower right region shown

in Figure 4. Dashed lines in red depict the novel links made

by Watts and Strogatz in 1998. The references connected by

the new links are labeled.

Table 1 summarizes the results of ﬁve NB regression mod-

els with different types of networks. They have an average

dispersion parameter θof 0.5270, which is equivalent to an α

of 1.8975. Coauthors has an average IRR of 1.3278, Ref-

erences has an average IRR of 1.0126, and Pages has an

average IRR of 0.9714. The effects of the three variables are

consistent and stable across the ﬁve types of networks. In con-

trast, the effects of structural variations are less stable. On the

other hand, structural variations appear to have a stronger

impact on global citations than do other more commonly

studied measures such as Coauthors and References.For

example, CL has an IRR of 3.160 in networks of co-cited refer-

ences and an IRR of 1.33 ×108in networks of noun phrases.

IRRs that are greater than 1.0 predict an increase of global

citations.

In theory, ZINB regression models may more accurately

estimate the effects of variables, especially in situations where

an excessive number of zeros is expected. It is widely known

that a considerable number of scientiﬁc publications would

never be cited. Table 2 summarizes a ZINB model derived

from networks of co-cited references (Model 1) and a corre-

sponding NB model (Model 2). The Vuong test indicated that

the ZINB model is superior to the NB model at p=0.0033.

The IRRs of References and Pages remain identical in

both models. The IRRs of Coauthors, MCR, and CL are

slightly lower in the ZINB model than those in the NB

model. Two variables have statistically signiﬁcant effects in

the zero-inﬂation model: MCR and References. In summary,

for this particular dataset, CL is a more prominent predictor

of global citations than are all other more commonly stud-

ied factors. Considering the fact that CL is associated with

a clearly deﬁned boundary-spanning mechanism, this ﬁnd-

ing provides empirical evidence that such mechanisms are

likely to reveal some fundamental insights into the creation

of scientiﬁc knowledge.

•Mass Extinction (1991–2010)

The Vuong test was not able to reject the hypothesis that

the ZINB model and the NB model are indistinguishable at

p=0.0928. In particular, none of the variables have statisti-

cally signiﬁcant effects on predicting the excessive number

of zeros (Table 3).

The ZINB/NB models reveal a similar pattern as that

seen in the Complex Network Analysis dataset. Again, CL is

the most prominent predictor of global citation counts, with

the strongest IRR coefﬁcient of 3.975 in the ZINB model.

443

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

FIG. 5. A close-up view of the lower right region shown in Figure 4, showing novel links between distinct clusters of co-cited references. [Color ﬁgure can

be viewed in the online version, which is available at wiley.com.]

TABLE 1. Negative binomial regression (NB) models of complex network analysis (1996–2004) at ﬁve different levels of granularity of units of

analysis.a

Data source: Complex network analysis (1996–2004), top-100 records per time slice, 2-year sliding window

Unit of analysis Reference Keyword Noun phrase Author Journal

Relation Co-citation Co-occurrence Co-occurrence Co-citation Co-citation

Offset (exposure) log2(Year) log2(Year) log2(Year) log2(Year) log2(Year)

No. of citing articles 3,515 3,072 3,254 3,271 3,271

Global citations Incidence rate ratios in NB models

Coauthors 1.306 0.000 1.298 0.000 1.326 0.000 1.359 0.000 1.350 0.000

Modularity change rate 1.083 0.025 1.038 0.086 1.047 0.305 1.055 0.276 1.060 0.180

Weighted cluster linkage 3.160 0.000 0.205 0.095 1.33 ×1080.000 2.879 0.000 1.204 0.049

Centrality Divergence 0.343 0.184 3.679 0.023 1.534 0.665 23.400 0.000 7.620 0.000

No. of references 1.013 0.000 1.013 0.000 1.013 0.000 1.012 0.000 1.012 0.000

No. of pages 0.970 0.000 0.971 0.000 0.971 0.000 0.973 0.000 0.972 0.000

Dispersion parameter (θ) 0.5284 0.5258 0.5150 0.5282 0.5375

2×log-likelihood −31,771 −28,331 −29,491 −29,506 −29,613

AIC 31,787 28,347 29,508 29,522 29,629

AIC =Akaike’s Information Criterion.

aReferences involve the least amount of ambiguity with the ﬁnest granularity whereas the other four types of units introduce ambiguity at various levels.

Models constructed with units of higher ambiguity are slightly improved in terms of AIC.

444 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

TABLE 2. Zero-inﬂated negative binomial regression (ZINB) and negative binomial regression (NB) models of global citation counts of 3,515 citing articles

on complex network analysis (1996–2004) with reference to networks of co-cited references using log2(year of publication) to account for the exposure

time.a

ZINB

Global cites Count model Negbin with log link Zero-inﬂation model Binomial with logit link NB

Coauthors 1.293 0.000 0.062 0.077 1.306 0.000

Modularity change rate 1.080 0.014 0.012 0.044 1.083 0.025

Weighted cluster linkage 3.103 0.000 1.304 0.906 3.160 0.000

Centrality Divergence 0.391 0.237 385363.6 0.102 0.343 0.184

No. of references 1.013 0.000 0.489 0.027 1.013 0.000

No. of pages 0.970 0.000 1.133 0.120 0.970 0.000

Dispersion parameter (θ) 0.536 0.528

AIC 31,768 31,787

Vuong test (ZINB>NB) −2.7186, p=0.0033

AIC =Akaike’s Information Criterion.

aCoefﬁcients are incidence rate ratios. Weighted cluster linkage is the strongest predictor of citation counts, followed by the number of coauthors, and the

modularity change rate. In this case, with a lower AIC and a statistically signiﬁcant Vuong test, the ZINB model is superior.

TABLE 3. Zero-inﬂated negative binomial regression (ZINB) and negative binomial regression (NB) models of global citation counts of 1,745 articles on

Mass Extinctions (1991–2010) with reference to networks of co-cited references using log2(year of publication) to account for the exposure time.a

ZINB

Global cites Count model Negbin with log link Zero-inﬂation model Binomial with logit link NB

Coauthors 1.160 0.001 0.191 0.380 1.162 0.0006

Modularity change rate 0.713 0.021 9.95 ×10−90.269 0.736 0.0374

Weighted cluster linkage 3.975 0.000 2.63 ×1050.461 3.805 0.0000

Centrality Divergence 9.837 0.334 3.65 ×10−20 0.877 8.619 0.3607

No. of references 1.004 0.000 1.025 0.283 1.004 0.0000

No. of pages 0.987 0.002 0.657 0.116 0.988 0.0048

Dispersion parameter (θ) 0.489 0.484

AIC 13,368 13,366

Vuong test (ZINB>NB) −1.3234, p=0.0928

AIC =Akaike’s Information Criterion.

aCoefﬁcients are incidence rate ratios. Cluster linkage is the strongest predictor of citation counts, followed by the number of coauthors and the number

of references, respectively. In this case, the ZINB and NB models are indistinguishable statistically.

The IRRs of Coauthors and References are slightly greater

than 1.0. The IRRs of MCR and Pages are less than 1.0, indi-

cating an expected decrease of citations as the two variables

increase independently.

•Terrorism (1996–2005)

Unlike the previous cases in which 1-year time slices

were used, we used time slices with longer exposure time

for the Terrorism (1996–2005) dataset, including 2- and 3-

year time slices. A 3-year moving window was used with

3-year time slices.

The ZINB and NB models revealed similar patterns to other

cases we have seen so far, except the strong IRR of Centrality

Divergence, which was absent in the earlier cases. The effects

of Coauthors,CL,References, and Pages are within the sim-

ilar range as shown in earlier cases. Table 4 includes models

for both 2- and 3-year time slices. The 3-year models have a

better relative goodness of ﬁt than the 2-year models accord-

ing to the AIC scores. The IRR of Centrality Divergence is

48.061 in 3-year slice models and 62.637 in 2-year slice mod-

els. We inspected articles with high Centrality Divergence to

ﬁnd an explanation of the strong Centrality Divergence.

Figure 6 illustrates connections made by the top-ﬁve arti-

cles with the strongest Centrality Divergence scores based

on 2-year time slice models. As show in the ﬁgure, much of

the boundary-spanning activities introduced by this group of

articles involve three major clusters: Cluster 12 biological ter-

rorism, Cluster 13 ocular injury, and Cluster 5 September 11.

The ﬁve top-ranked high-Centrality Divergence articles are

listed in Table 5, including their DOIs. One can tell from

their titles which topic areas that they were connecting; for

example, connections between terrorism and medical services

(Papers 1 and 3) and connections between PTSD and Septem-

ber 11 terrorist attacks (Papers 2, 4, and 5). In terms of the

clusters shown in Figure 6, the topic of bioterrorism is clearly

associated with Cluster 12, the topic of September 11 terrorist

attacks is the focus of Cluster 5, and the topic of clinical and

medical services is closely related to Cluster 13.

The current analysis suggests that the strong IRR of

Centrality Divergence may be due to the wide-ranging inter-

disciplinary structure of the subject matter. If this is indeed

the case, then a strong Centrality Divergence can be a valu-

able early sign of transformative research at interdisciplinary

445

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

TABLE 4. Zero-inﬂated negative binomial regression (ZINB) and negative binomial regression (NB) models of global citation counts of 3,476 articles on

Terrorism (1996–2005) with reference to networks of co-cited references using log2(year of publication) to account for the exposure time.a

ZINB

Global cites Count model Negbin with log link Zero-inﬂation model Binomial with logit link NB

Slice length =3 years

Coauthors 1.808 0.0000 0.000 0.9524 1.960 0.0000

Modularity change rate 0.578 0.0973 1.563 0.8713 0.581 0.0603

Weighted cluster linkage 3.306 0.0001 0.000 0.7251 3.300 0.0002

Centrality Divergence 48.061 0.0012 0.121 0.9368 64.120 0.0000

No. of references 1.018 0.0000 0.980 0.0702 1.019 0.0000

No. of pages 0.985 0.0000 1.002 0.8916 0.986 0.0048

Dispersion parameter (θ) 0.520 0.464

AIC 20,591 20,619

Vuong test (ZINB>NB) −2.9912, p=0.0014

Slice length =2 years

Coauthors 1.853 0.0000 0.067 0.0119 2.005 0.0000

Modularity change rate 0.586 0.0000 0.394 0.3708 0.596 0.0000

Weighted cluster linkage 3.180 0.0000 28.363 0.0435 2.674 0.0000

Centrality Divergence 62.637 0.0319 0.000 0.2143 108.959 0.0046

No. of references 1.018 0.0000 0.984 0.1296 1.019 0.0000

No. of pages 0.986 0.0000 1.004 0.6902 0.986 0.0000

Dispersion parameter (θ) 0.518 0.464

AIC 20,668 20,695

Vuong test (ZINB>NB) −2.8463, p=0.0022

AIC =Akaike’s Information Criterion.

aThe length of each time slice and the width of the moving window are both 3 years. Coefﬁcients are IRRs.The strongest predictor is Centrality Divergence,

followed by weighted cluster linkage, the number of coauthors, and the number of references. With a lower AIC and a statistically signiﬁcant Vuong test, the

ZINB model is a better model than is the NB model.

FIG. 6. Co-citations made by the ﬁve articles with the strongest Centrality Divergence scores in Terrorism (1996–2005) that reinforced existing patterns

(left) and introduced novel connections (right). Structural variations were computed based on 2-year time slices. [Color ﬁgure can be viewed in the online

version, which is available at wiley.com.]

levels. Further research is needed to gain additional insights

into implications of these new structural variation metrics.

•CiteSpace Expanded (2001–2010)

This dataset was formed with a seed article (Chen, 2006)

by including papers that have at least one cited reference in

common with the seed article during the period of 2001 and

2010. Estimates made by ZINB and NB models are listed in

Table 6.TheVuong test was not able to reject the null hypoth-

esis that the ZINB and NB models are indistinguishable.

A ZINB regression model shows a statistically signiﬁcant

446 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

TABLE 5. The top-ﬁve articles in Terrorism (1996–2005) with the strongest Centrality Divergence scores. Co-citation connections made by these papers

are shown in Figure 6.

No. Cites CKL Author Year Title. Source. DOI

1 23 0.500 Flowers, L.K. 2002 Bioterrorism preparedness II: The community and emergency medical services systems. Emergency

Medicine Clinics of North America.10.1016/S0733-8627(01)00009-8

2 35 0.327 Grieger, T.A. 2003 Posttraumatic stress disorder alcohol use and perceived safety after the terrorist attack on the Pentagon.

Psychiatric Services. 10.1176/appi.ps.54.10.1380

3 72 0.285 Stein, M. 1999 Medical consequences of terrorism—The conventional weapon threat. Surgical Clinics of North America.

10.1016/S0039-6109(05)70091-8

4 80 0.271 Ahern, J. 2002 Television images and psychological symptoms after the September 11 terrorist attacks. Psychiatry.

10.1521/psyc.65.4.289.20240

5 23 0.264 Ford, C.A. 2003 Reactions of young adults to September 11 2001. Archives of Pediatrics & Adolescent Medicine.

10.1001/archpedi.157.6.572

TABLE 6. Zero-inﬂated negative binomial regression (ZINB) and negative binomial regression (NB) models of global citation counts of 3,260 articles

published between 2001 and 2010 in an area seeded by our 2006 Journal of the American Society of Information Science and Technology article with reference

to networks of co-cited references using log2(year of publication) to offset the exposure time.a

ZINB

Global cites Count model Negbin with log link Zero-inﬂation model Binomial with logit link NB

Coauthors 1.070 0.0826 4.125 0.3379 1.084 0.0225

Modularity change rate 1.025 0.5482 1.715 0.0198 1.013 0.7135

Weighted cluster linkage 3.230 0.0046 0.001 0.0834 3.589 0.0000

Centrality Divergence 3.203 0.0870 0.000 0.1590 3.453 0.0272

No. of references 1.010 0.0000 0.958 0.0124 1.010 0.0000

No. of pages 0.985 0.0000 1.097 0.0844 0.984 0.0048

Dispersion parameter (θ) 0.423 0.413

AIC 19,797 19,786

Vuong test (ZINB>NB) −0.4694, p=0.3194

AIC =Akaike’s Information Criterion.

aWeighted cluster linkage is the strongest predictor of citation counts, followed by the number of references. In this case, the ZINB and NB models are

not indistinguishable statistically.

TABLE 7. The Akaike’s Information Criterion (AIC) of zero-inﬂated negative binomial regression (ZINB) and negative binomial regression (NB) models.

ZINB models are distinguishable from NB models in three of the ﬁve cases. AICs increase as the number of papers per time slice increases.

Cases Seeded From Years nn/Year AIC (ZINB) AIC (NB) ZINB >NB (p)

Mass Extinctions No 1991–2010 20 1,745 87.25 13,368 13,366 0.0928

CiteSpace Expanded Yes 2001–2010 10 3,260 163.00 19,797 19,786 0.3194

Terrorism (3-year slices) No 1996–2005 10 3,476 173.80 20,591 20,619 0.0014

Terrorism (2-year slices) No 1996–2005 10 3,476 173.80 20,668 20,695 0.0022

Complex Networks Yes 1996–2004 9 3,515 175.75 31,768 31,787 0.0033

main effect of the CL variation with an IRR of 3.230. Ref-

erences and Pages have similar values as those in other cases.

The NB model reveals a strong effectof Centrality Divergence

(IRR =3.453).

The seed article describes CiteSpace, a visual analytic tool

for identifying emerging trends and changes in scientiﬁc lit-

erature. It has been cited 100 times in the Web of Science and

has the highest CL score (12.78) and the highest Centrality

Divergence score (1.24).

Discussion

ZINB models are superior to their NB counterparts in the

three cases we studied, but indistinguishable in the other two

cases (see Table 7). The AIC values of these models increase

as the number of citing articles per time slices increases,

but there is no apparent trend in connection to whether

the datasets are seeded. Mass Extinctions and CiteSpace

Expanded have the lowestAIC values whereas Complex Net-

works has the highest AIC value. In the Terrorism case, the

use of 3-year time slices improved the models over the 2-

year conﬁgurations. These results seem to suggest that ZINB

models are particularly recommended as the size of a dataset

increases. As the number of articles increases, more and more

zero-citation articles need to be accounted for.

The IRR of a variable has an intuitive interpretation. If it

is greater than 1.0 and all other variables are held constant,

447

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

TABLE 8. Incidence rate ratios (IRRs) of predictors of global citation counts.

Cases Coauthors MCR Cluster linkage Centrality Divergence References Length in pages

Mass Extinctions 1.160a0.713 3.975 9.837 1.004 0.987

CiteSpace Expanded 1.084 1.013 3.589 3.453 1.010 0.984

Terrorism (3-year slices) 1.808 0.578 3.306 48.061 1.018 0.985

Terrorism (2-year slices) 1.853 0.586 3.180 62.637 1.018 0.986

Complex Networks 1.293 1.080 3.103 0.391 1.013 0.970

M1.440 0.794 3.431 24.876 1.013 0.982

MCR =modularity change rate.

aThe IRRs in bold are statistically signiﬁcant at p=0.05. The IRRs of coauthors, cluster linkage, references, and pages are consistent across the cases

whereas the IRRs of MCR and Centrality Divergence are mixed.

then the citation count is expected to increase along with a

unit increase of the variable. In contrast, if it is less than 1.0,

then the citation count is expected to decrease as the variable

increases. Cluster Linkage,Coauthors, and References are

persistently found to predict an increase of citation counts.

The strongest predictor, Cluster Linkage, has an average IRR

of 3.431, which is more than twice that of the second strongest

predictor, Coauthors (average IRR =1.440) and more than

three times that of the third strongest predictor, References

(average IRR =1.013).The length of an article in terms of the

number of pages has a minor, but consistent, negative impact

with an average IRR of 0.982 across all ﬁve cases.

In contrast, the IRR estimates of Centrality Diver-

gence and MCR did not give clear signals. Although the

largest average IRR (24.876) is found with Centrality Diver-

gence and although in three of the ﬁve cases the effect was

positive and statistically signiﬁcant, it has a negative impact

on citations in the Complex Network case. It is possible

that Centrality Divergence is sensitive to the structure of

the underlying network as well as to the structural change

of the network. Given its strong IRRs in its magnitude and

the possible sensitivity, the behavior of the metric should

be investigated further to explain the discrepancies across

different cases.

MCR is an indirect measure of the impact of cross-cluster

links. It has an average IRR of 0.794, but the estimated IRRs

are inconsistent across the ﬁve cases in terms of statistical

signiﬁcance and the direction of citation change. The dis-

crepancies may reﬂect the connectivity of individual nodes

involved. For example, adding a new link to a node with either

a large degree or a small degree is likely to affect the MCR

much more than is linking two nodes of similar connectivity.

Cluster Linkage is a more direct measure of intellectual

potential than are Coauthors and References.Cluster Linkage

has a simple and clear theoretical interpretation supported by

the underlying boundary-spanning mechanism. The strong

IRRs of Cluster Linkage suggest that the structural variation

metric is more efﬁcient for predicting the growth of citation

counts than are indicators such as Coauthors and References.

These ﬁndings are very encouraging because of their

theoretical and practical implications. These ﬁndings are

valuable in improving our understanding of how transfor-

mative ideas can be made and recognized. The premise of

the structural variation approach is that transformative ideas

are expected to introduce signiﬁcant and computationally

detectable structural changes. The value of such approaches

has been identiﬁed by many researchers (Boyack et al., 2005;

Leydesdorff, 2001; Shibata et al., 2007; Upham et al., 2010;

van Dalen & Kenkens, 2005). The boundary-spanning mech-

anism provides one possible explanation of what factors

attract citations. It is conceivable that an idea has the potential

to introduce novel structural changes, but fails to material-

ize it due to many reasons. In the present study, we use the

co-citation network immediately prior to the publication of

a new paper as the reference point. There are several other

options at various levels of granularity and scale. For instance,

in the Complex Network Analysis case, structural variation

models based on networks of co-occurring keywords have the

best AIC score with an IRR of Centrality Divergence of 3.679.

Models based on networks of co-cited authors led to a strong

Centrality Divergence IRR of 23.400. We did not investigate

the implications of these results in the present article, but we

believe that they do indicate a wide range of potentially sig-

niﬁcant directions of research that is far beyond the scope of

a single study.

Conclusions

We have found statistical evidence of the boundary-

spanning mechanism. An article that introduces novel con-

nections between clusters of co-cited references is likely to

subsequently become highly cited. In addition, we have found

that the IRRs of Cluster Linkage are more than twice as much

as the IRRs of Coauthors and References. This ﬁnding pro-

vides a more fundamental explanation of why the number of

references cited by an article appears to be a good predictor

of its future citations, as found in many previous studies.As a

result, the structural variation paradigm clariﬁes why a num-

ber of extrinsic features appear to be associated with high

citations.

A distinct characteristic of the structural variation

approach is the focus on the potential connection between

the degree of structural variation introduced by an article

and its future impact. The analytic and modeling procedure

demonstrated in this article is expected to serve as an exem-

plar for subsequent studies along this line of research. More

448 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

important, the focus on the underlying mechanisms of sci-

entiﬁc activity is expected to provide additional insights and

practical guidance for scientists, sociologists, historians, and

philosophers of scientiﬁc knowledge.

There are many new challenges and opportunities ahead.

For example, how common is the boundary-spanning mech-

anism in scientiﬁc discoveries overall? What are the other

major mechanisms, and how do they interact with the

boundary-spanning mechanism? There are other potentially

valuable techniques that we have not utilized in the present

study, including topic modeling, citation context analysis,

survival analysis, and burst detection. In short, a lot of work

needs to be done, and this is an encouraging start.

We conclude that structural variation is an essential aspect

of the development of scientiﬁc knowledge and it has the

potential to reveal the underlying mechanisms of the growth

of scientiﬁc knowledge. The focus on the underlying mecha-

nisms of knowledge creation is the key to the predictive poten-

tial of the structural variation approach. The theory-driven

explanatory and computational approach sets an extensible

framework for detecting and tracking potentially creative

ideas and gaining insights into challenges and opportunities

in light of the collective wisdom.

Acknowledgments

I thank the anonymous reviewers for their valuable

comments.

References

Aksnes, D.W. (2003). Characteristics of highly cited papers. Research

Evaluation, 12(3), 159–170.

Barabási, A.L., & Albert, R. (1999). Emergence of scaling in random

networks. Science, 286(5439), 509–512.

Bornmann, L., & Daniel, H.-D. (2006). What do citation counts measure?

A review of studies on citing behavior. Journal of Documentation, 64(1),

45–80.

Boyack, K.W., Klavans, R., Ingwersen, P., & Larsen, B. (2005, July). Pre-

dicting the importance of current papers. Paper presented at the 10th

International Conference of the International Society for Scientometrics

and Informetrics, Stockholm, Sweden. Retrieved from https://cfwebprod

.sandia.gov/cfdocs/CCIM/docs/kwb_rk_ISSI05b.pdf

Brody, T., & Harnad, S. (2005). Earlier web usage statistics as predictors of

later citation impact. Retrieved from http://arxiv.org/ftp/cs/papers/0503/

0503020.pdf

Burt, R.S. (2004). Structural holes and good ideas. American Journal of

Sociology, 110(2), 349–399.

Buter, R., Noyons, E., & van Raan, A. (2011). Searching for converging

research using ﬁeld to ﬁeld citations. Scientometrics, 86(2), 325–338.

Chen, C. (2003). Mapping scientiﬁc frontiers: The quest for knowledge

visualization. London: Springer-Verlag.

Chen, C. (2006). CiteSpace II: Detecting and visualizing emerging trends and

transient patterns in scientiﬁc literature. Journal of the American Society

for Information Science and Technology, 57(3), 359–377.

Chen, C. (2011). Turning points: The nature of creativity. New York:

Springer.

Chen, C., Chen, Y., Horowitz, M., Hou, H., Liu, Z., & Pellegrino, D. (2009).

Towards an explanatory and computational theory of scientiﬁc discovery.

Journal of Informetrics, 3(3), 191–209.

Chen, C., Cribbin, T., Macredie, R., & Morar, S. (2002). Visualizing and

tracking the growth of competing paradigms: Two case studies. Journal

of the American Society for Information Science and Technology, 53(8),

678–689.

Chen, C., Ibekwe-SanJuan, F., & Hou, J. (2010).The structure and dynamics

of co-citation clusters:A multiple-perspective co-citation analysis. Journal

of the American Society for Information Science and Technology, 61(7),

1386–1409.

Chen, C., Lin, X., & Zhu, W. (2006, November). Trailblazing through a

knowledge space of science: Forwardcitation expansion in CiteSeer. Paper

presented at the 69th annual meeting of the American Society for Infor-

mation Science and Technology (ASIS&T 2006), Austin, TX. Retrieved

from http://eprints.rclis.org/archive/00008019/

Chubin, D.E. (1994 ). Grants peer-review in theory and practice. Evaluation

Review, 18(1), 20–30.

Chubin, D.E., & Hackett, E.J. (1990). Paperless science: Peer review and

U.S. science policy. Albany, NY: State University of New York Press.

Cuhls, K. (2001). Foresight with Delphi surveys in Japan. Technology

Analysis & Strategic Management, 13(4), 555–569.

Dewett, T., & Denisi, A.S. (2004). Exploring scholarly reputation: It’s more

than just productivity. Scientometrics, 60(2), 249–272.

Fauconnier, G., & Turner, M. (1998). Conceptual integration networks.

Cognitive Science, 22(2), 133–187.

Fleming, L., & Bromiley, P. (2000, December). A variable risk propen-

sity model of technological risk taking. Paper presented at the Applied

Statistics Workshop. Retrieved from http://courses.gov.harvard.edu/

gov3009/fall00/ﬂeming.pdf

Galea, S., Ahern, J., Resnick, H., Kilpatrick, D., Bucuvalas, M., Gold, J., &

Vlahov, D. (2002). Psychological sequelae of the September 11 terrorist

attacks in New York City. New England Journal of Medicine, 346(13),

982–987.

Garﬁeld, E. (1955). Citation indexes for science: A new dimension in

documentation through association of ideas. Science, 122, 108–111.

Haslam, N., Ban, L., Kaufmann, L., Loughnan, S., Peters, K., Whelan, J., &

Wilson, S. (2008). What makes an article inﬂuential? Predicting impact

in social and personality psychology. Scientometrics, 76(1), 169–185.

Häyrynen, M. (2007). Breakthrough research: Funding for high-risk research

at the Academy of Finland. Helsinki: Academy of Finland.

Hettich, S., & Pazzani, M.J. (2006). Mining for proposal reviewers: Lessons

learned at the National Science Foundation. In Proceedings of the 12th

ACM SIGKDD International Conference on Knowledge Discovery and

Data Mining (pp. 862–871). New York: ACM Press.

Hilbe, J.M. (2011). Negative binomial regression (2nd ed.). New York:

Cambridge University Press.

Hirsch, J.E. (2007). Does the h index have predictive power? Proceedings

of the National Academy of Sciences, USA, 104(49), 19193–19198.

Hsieh, C. (2011). Explicitly searching for useful inventions: Dynamic relat-

edness and the costs of connecting versus synthesizing. Scientometrics,

86(2), 381–404.

Kostoff, R. (2007). The difference between highly and poorly cited medical

articles in the journal Lancet. Scientometrics, 72, 513–520.

Kurtz, M.J., Eichhorn, G., Accomazzi, A., Grant, C., Demleitner, M.,

Henneken, E., & Murray, S.S. (2005). The effect of use and access on

citations. Information Processing & Management, 41(6), 1395–1402.

Lahiri, M., Maiya, A.S., Sulo, R., Habiba, & Wolf, T.Y.B. (2008,

December). The impact of structural changes on predictions of diffusion

in networks. Paper presented at the the 2008 IEEE International Con-

ference on Data Mining Workshops (ICDMW ’08), Pisa, Italy. Retrieved

from http://compbio.cs.uic.edu/∼mayank/papers/LahiriMaiyaSuloHabiba

BergerWolf_ImpactOfStructuralChanges08.pdf

Lambert, D. (1992). Zero-infated Poisson regression, with an application to

defects in manufacturing. Technometrics, 34, 1–14.

Levitt, J., & Thelwall, M. (2008). Patterns of annual citation of highly cited

articles and the prediction of their citation ranking: A comparison across

subjects. Scientometrics, 77(1), 41–60.

Leydesdorff, L. (2001). The challenge of scientometrics: The development,

measurement, and self-organization of scientiﬁc communications. Boca

Raton, FL: Universal Publishers.

Lipinski, C., & Hopkins, A. (2004). Navigating chemical space for biology

and medicine. Nature, 432(7019), 855–861.

449

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY—March 2012

DOI: 10.1002/asi

Luxburg, U. von. (2006). A tutorial on spectral clustering. Retrieved

from http://www.kyb.mpg.de/ﬁleadmin/user_upload/ﬁles/publications/

attachments/Luxburg07_tutorial_4488%5b0%5d.pdf

Martin, B.R. (2010). The origins of the concept of “foresight” in science

and technology: An insider’s perspective. Technological Forecasting and

Social Change, 77(9), 1438–1447.

Merton, R.K. (1968). The Mathew Effect in science. Science, 159(3810),

56–63.

Miles, I. (2010). The development of technology foresight: A review.

Technological Forecasting and Social Change, 77(9), 1448–1456.

Newman, M.E.J. (2006). Modularity and community structure in net-

works. Proceedings of the National Academy of Sciences, USA, 103(23),

8577–8582.

Persson, O. (2010). Are highly cited papers more international? Scientomet-

rics, 83(2), 397–401.

Price, D.D. (1965). Networks of scientiﬁc papers. Science, 149, 510–515.

Shibata, N., Kajikawa, Y., & Matsushima, K. (2007). Topological analy-

sis of citation networks to discover the future core articles. Journal of

the American Society for Information Science and Technology, 58(6),

872–882.

Skilton, P. (2009). Does the human capital of teams of natural science authors

predict citation frequency? Scientometrics, 78(3), 525–542.

Swanson, D.R. (1986a). Fish oil, Raynaud’s syndrome, and undiscovered

public knowledge. Perspectives in Biology and Medicine, 30, 7–18.

Swanson, D.R. (1986b). Undiscovered public knowledge. Library Quarterly,

56(2), 103–118.

Takeda,Y., & Kajikawa,Y. (2010). Tracking modularity in citation networks.

Scientometrics, 83(3), 783.

Tichy, G. (2004). The over-optimism among experts in assessment

and foresight. Technological Forecasting and Social Change, 71(4),

341–363.

Tijssen, R.J.W., Visser, M.S., & van Leeuwen, T.N. (2002). Benchmarking

international scientiﬁc excellence: Are highly cited research papers an

appropriate frame of reference? Scientometrics, 54(3), 381–397.

Upham, S.P., Rosenkopf, L., & Ungar, L.H. (2010). Positioning knowl-

edge: Schools of thought and new knowledge creation. Scientometrics, 83,

555–581.

van Dalen, H.P., & Kenkens, K. (2005). Signals in science: On the impor-

tance of signaling in gaining attention in science. Scientometrics, 64(2),

209–233.

Walters, G.D. (2006). Predicting subsequent citations to articles published in

twelve crime-psychology journals: Author impact versus journal impact.

Scientometrics, 69(3), 499–510.

Watts, D.J., & Strogatz, S.H. (1998). Collective dynamics of “small-world”

networks. Nature, 393(6684), 440–442.

Weeber, M. (2003). Advances in literature-based discovery. Journal of

the American Society for Information Science and Technology, 54(10),

913–925.

Zeileis, A., Kleiber, C., & Jackman, S. (2011). Regression models for count

data in R. Retrieved from http://cran.r-project.org/web/packages/pscl/

vignettes/countreg.pdf

Knowledge domain and emerging trends in multimorbidity and frailty research from 2003 to 2023: a scientometric study using citespace and VOSviewer

Article

Full-text available

Oct 2023

Background Multimorbidity and frailty represent emerging global health burdens that have garnered increased attention from researchers over the past two decades. We conducted a scientometric analysis of the scientific literature on the coexistence of multimorbidity and frailty to assess major research domains, trends, and inform future lines of research. Methods We systematically retrieved scientific publications on multimorbidity and frailty from the Web of Science Core Collection, spanning from 2003 to 2023. Scientometric analysis was performed using CiteSpace and VOSviewer, enabling the visualization and evaluation of networks comprising co-citation references, co-occurring keywords, countries, institutions, authors, and journals. Results A total of 584 eligible publications were included in the analysis. An exponential rise in research interest in multimorbidity and frailty was observed, with an average annual growth rate of 47.92% in publications between 2003 and 2022. Three major research trends were identified: standardized definition and measurement of multimorbidity and frailty, comprehensive geriatric assessment utilizing multimorbidity and frailty instruments for older adults, and the multifaceted associations between these two conditions. The United States of America, Johns Hopkins University, Fried LP, and the Journal of the American Geriatrics Society were identified as the most influential entities within this field, representing the leading country, institution, author, and journal, respectively. Conclusions Scientometric analysis provides invaluable insights to clinicians and researchers involved in multimorbidity and frailty research by identifying intellectual bases and research trends. While the instruments and assessments of multimorbidity and frailty with scientific validity and reliability are of undeniable importance, further investigations are also warranted to unravel the underlying biological mechanisms of interactions between multimorbidity and frailty, explore the mental health aspects among older individuals with multimorbidity and frailty, and refine strategies to reduce prescriptions in this specific population.

A bibliometric analysis on discovering anti-quorum sensing agents against clinically relevant pathogens: current status, development, and future directions

Article

Full-text available

Nov 2023

Background Quorum sensing is bacteria’s ability to communicate and regulate their behavior based on population density. Anti-quorum sensing agents (anti-QSA) is promising strategy to treat resistant infections, as well as reduce selective pressure that leads to antibiotic resistance of clinically relevant pathogens. This study analyzes the output, hotspots, and trends of research in the field of anti-QSA against clinically relevant pathogens. Methods The literature on anti-QSA from the Web of Science Core Collection database was retrieved and analyzed. Tools such as CiteSpace and Alluvial Generator were used to visualize and interpret the data. Results From 1998 to 2023, the number of publications related to anti-QAS research increased rapidly, with a total of 1,743 articles and reviews published in 558 journals. The United States was the largest contributor and the most influential country, with an H-index of 88, higher than other countries. Williams was the most productive author, and Hoiby N was the most cited author. Frontiers in Microbiology was the most prolific and the most cited journal. Burst detection indicated that the main frontier disciplines shifted from MICROBIOLOGY, CLINICAL, MOLECULAR BIOLOGY, and other biomedicine-related fields to FOOD, MATERIALS, NATURAL PRODUCTS, and MULTIDISCIPLINARY. In the whole research history, the strongest burst keyword was cystic-fibrosis patients, and the strongest burst reference was Lee and Zhang (2015). In the latest period (burst until 2023), the strongest burst keyword was silver nanoparticle, and the strongest burst reference was Whiteley et al. (2017). The co-citation network revealed that the most important interest and research direction was anti-biofilm/anti-virulence drug development, and timeline analysis suggested that this direction is also the most active. The key concepts alluvial flow visualization revealed seven terms with the longest time span and lasting until now, namely Escherichia coli , virulence, Pseudomonas aeruginosa , virulence factor, bacterial biofilm, gene expression, quorum sensing. Comprehensive analysis shows that nanomaterials, marine natural products, and artificial intelligence (AI) may become hotspots in the future. Conclusion This bibliometric study reveals the current status and trends of anti-QSA research and may assist researchers in identifying hot topics and exploring new research directions.

Big data visualisation in regional comprehensive economic partnership: a systematic review

Article

Full-text available

Nov 2023

Lijun Li

The Regional Comprehensive Economic Partnership (RCEP) is an agreement that transformed the world economy and entered into force in January 2022 with the participation of fifteen nations. In the study, the visualisation analysis was 301 articles in Web of Science (WoS) on the subjects of “RCEP,” or “The Regional Comprehensive Economic Partnership,” from January 2012 to January 2023, using CiteSpace. The results of a comparative analysis of the number of journals co-citation and keyword co-occurrence indicate that further studies of “RCEP” will not be limited to the scope of traditional economics, but more and further fields are waiting for scholars to develop.

A decennial study of the trend of antibiotic studies in China

Article

Full-text available

Nov 2023
ENVIRON SCI POLLUT R

Antibiotics are one of the greatest inventions in human history and are used worldwide on an enormous scale. Besides its extensive usage in medical and veterinary arenas to treat and prevent the infection, its application is very prominent in other fields, including agriculture, aquaculture, and horticulture. In recent decades, the increased consumption of antibiotics in China saw a vast increase in its production and disposal in various environments. However, in this post-antibiotic era, the abuse and misuse of these valuable compounds could lead to the unreversible consequence of drug resistance. In China, antibiotics are given a broad discussion in various fields to reveal their impact on both human/animals health and the environment. To our knowledge, we are the first paper to look back at the development trend of antibiotic-related studies in China with qualitative and quantitative bibliometric analysis from the past decades. Our study identified and analyzed 5559 papers from its inception (1991) to December 6, 2021, from the Web of Science Core Collection database. However, with few authors and institutions focusing on long-term studies, we found the quality of contributions was uneven. Studies mainly focused on areas such as food science, clinical research, and environmental studies, including molecular biology, genetics and environmental, ecotoxicology, and nutrition, which indicate possible primary future trends. Our study reports on including potentially new keywords, studies’ milestones, and their contribution to antibiotic research. We offer potential topics that may be important in upcoming years that could help guide future research.

Global trends and future prospects of lactic acid production from lignocellulosic biomass

Article

Full-text available

Nov 2023

Lignocellulosic biomass (LCB) stands as a substantial and sustainable resource capable of addressing energy and environmental challenges. This study employs bibliometric analysis to investigate research trends in lactic acid (LA) production from LCB spanning the years 1991 to 2022. The analysis reveals a consistent growth trajectory with minor fluctuations in LA production from LCB. Notably, there's a significant upswing in publications since 2009. Bioresource Technology and Applied Microbiology and Biotechnology emerge as the top two journals with extensive contributions in the realm of LA production from LCB. China takes a prominent position in this research domain, boasting the highest total publication count (736), betweenness centrality value (0.30), and the number of collaborating countries (42), surpassing the USA and Japan by a considerable margin. The author keywords analysis provides valuable insights into the core themes in LA production from LCB. Furthermore, co-citation reference analysis delineates four principal domains related to LA production from LCB, with three associated with microbial conversion and one focused on chemical catalytic conversion. Additionally, this study examines commonly used LCB, microbial LA producers, and compares microbial fermentation to chemical catalytic conversion for LCB-based LA production, providing comprehensive insights into the current state of this field and suggesting future research directions.

A Scientometirc Analysis on Translation Studies of Tao Te Ching in China Based on CiteSpace

Article

Full-text available

Oct 2023

Xiujuan Zhang

Based on the data of research papers published between 1988 to 2023, which were retrieved from China National Knowledge Infrastructure (CNKI), this paper makes a scientometric analysis of the research focuses of English translations studies of Tao Te Ching in China by employing CiteSpace. The findings were presented in knowledge domains. Aiming to uncover the research development and hot-discussed topics, the paper probes into the number of published articles, the high-impact authors and institutions, high-frequency keywords. It is found that translator research and translation strategies are the focuses of this research field. Current research mainly focuses on three aspects: translators such as Roger T. Ames, Lin Yutang, D. C. Lau, and Legge; translation strategies such as domestication and foreignization; English translation of key words such as Tao.

Bibliometric Analysis of Research Trends on Manual Therapy for Low Back Pain Over Past 2 Decades

Article

Full-text available

Sep 2023

Purpose Low back pain (LBP) is a prevalent musculoskeletal disorder, and manual therapy (MT) is frequently employed as a non-pharmacological treatment for LBP. This study aims to explore the research hotspots and trends in MT for LBP. MT has gained widespread acceptance in clinical practice due to its proven safety and effectiveness. The study aims to analyze the developments in the field of MT for LBP over the past 23 years, including leading countries, institutions, authoritative authors, journals, keywords, and references. It endeavors to provide a comprehensive summary of the existing research foundation and to analyze the current cutting-edge research trends. Methods Relevant articles between 2000 and 2023 were retrieved from the Web of Science Core Collection (WOSCC) database. We used the software VOSviewer and CiteSpace to perform the analysis and summarize current research hotspots and emerging trends. Results Through screening, we included 1643 papers from 2000 to 2023. In general, the number of articles published each year showed an upward trend. The United States had the highest number of publications and citations. Canadian Memorial Chiropractic College was the most published research institution. The University of Pittsburgh in the United States had the most collaboration with other research institutions. Long, Cynthia R. was the active author. Journal of Manipulative and Physiological Therapeutics was the most prolific journal with 234 publications. Conclusion This study provides an overview of the current status and trends of clinical studies on MT for LBP in the past 23 years using the visualization software, which may help researchers identify potential collaborators and collaborating institutions, hot topics, and new perspectives in research frontiers, while providing new clinical practice ideas for the treatment of LBP.

Publication Trends and Hot Topics in Dysplastic Nevus Research: A 30-Year Bibliometric Analysis

Article

Full-text available

Oct 2023

Hazal İzol Özmen

Introduction Dysplastic nevi are pigmented lesions that exhibit clinical and histological features of both common nevi and melanoma. In recent years, there has been an increase in publications on dysplastic nevi. Bibliometric analysis is a method of evaluating trends in large number of publications and identifying popular topics. Objectives The objective of this study is to provide an overview of the landscape of publications related to dysplastic nevi, visualize trends and identify popular topics in the literature. Methods Thomson Reuters’ Web of Science database was searched with the following query in title, abstract or keywords: TS = (“dysplastic nevus” OR “clark nevus” OR “atypical nevus” OR “dysplastic nevi” OR “clark nevi” OR “atypical nevi”). Time span was set to 1992–2022. Document type was set to Article. Titles, authors, abstracts, institutions, countries, journals, references, and the citation information were recorded. Results Although the number of publications has declined over time, the USA remains the leading contributor to published articles. Key clusters of frequently used keywords were identified. The Journal of the American Academy of Dermatology had the highest number of published titles. Country and journal analysis were supplemented by co-citation and co-cited reference cluster analysis. Burst analyses revealed authors like Kittler, Argenziano, and Gandini as significant contributors, with their works receiving strong citation bursts extending until the end of the study period. Conclusions This bibliometric analysis revealed trends and interest pockets in the literature pertaining to dysplastic nevi and melanoma. This study aids in understanding the current research landscape and highlights potential future directions in this field.

Research Hotspot and Front Visualization of the Shadow Education System: Data from Web of Science

Article

Oct 2023

This scientometric review takes 351 documents from 1992–2021 as the research object based on the Web of Science database. With the help of CiteSpace, this study aims to construct visualization mapping knowledge domains, display the research status in shadow education more intuitively, contribute opportunities for further research, and provide a more visual basis for dialog among researchers, policymakers and interested actors in the field. This study, by building coauthor, coword, and cocitation knowledge visualization maps, demonstrates cooperation among authors, research hotspots and frontiers in the field. Our results show that shadow education has experienced a rapid expansion over the last decade but that the scope of the collaborative circle of academia needs to be further expanded. Furthermore, because of shadow education’s variable forms, researchers need to pay extra attention to the scope of its definition. Parents are involved in too many of their children’s educational choices; learning requires more self-drive and improved self-learning ability.

Tourism Carbon Emissions: A Systematic Review of Research Based on Bibliometric Methods

Article

Oct 2023

Trailblazing through a Knowledge Space of Science: Forward Citation Expansion in CiteSeer

Conference Paper

Oct 2007

DOI: https://doi.org/10.1002/meet.1450430180

Conceptual integration networks

Chapter

Aug 2008

Zero-Inflated Poisson Regression, With an Application to Defects in Manufacturing

Article

Jan 1992

D Lambert

Mapping Scientific Frontiers: The Quest for Knowledge Visualization

Book

Jan 2003

Chaomei Chen

This is an examination of the history and the state of the art of the quest for visualizing scientific knowledge and the dynamics of its development. Through an interdisciplinary perspective this book presents profound visions, pivotal advances, and insightful contributions made by generations of researchers and professionals, which portrays a holistic view of the underlying principles and mechanisms of the development of science. This updated and extended second edition:highlights the latest advances in mapping scientific frontiersexamines the foundations of strategies, principles, and design patternsprovides an integrated and holistic account of major developments across disciplinary boundariesAnyone who tries to follow the exponential growth of the literature on citation analysis and scientometrics knows how difficult it is to keep pace. Chaomei Chen has identified the significant methods and applications in visual graphics and made them clear to the uninitiated. Derek Price would have loved this book which not only pays homage to him but also to the key players in information science and a wide variety of others in the sociology and history of science. Eugene GarfieldThis is a wide ranging book on information visualization, with a specific focus on science mapping. Science mapping is still in its infancy and many intellectual challenges remain to be investigated and many of which are outlined in the final chapter. In this new edition Chaomei Chen has provided an essential text, useful both as a primer for new entrants and as a comprehensive overview of recent developments for the seasoned practitioner. Henry SmallChaomei Chen is a Professor in the College of Information Science and Technology at Drexel University, Philadelphia, USA, and a ChangJiang Scholar at Dalian University of Technology, Dalian, China. He is the Editor-in-Chief of Information Visualization and the author of Turning Points: The Nature of Creativity (Springer, 2012) and Information Visualization: Beyond the Horizon (Springer, 2004, 2006).

A tutorial on spectral clustering

Article

Jan 2007
STAT COMPUT

U. Von Luxburg

Tracking modularity in citation networks

Article

Jan 2009

Clustering using cocitation and bibliographic coupling is an effective tool to analyze the structure of scientific research, but its arbitrariness in the setting of a clustering threshold is a problem. This study tracked modularity of citation networks in research domains, and we found that there are three stages in clustering of citation networks and it is universal across our case studies. In the first stage, core clusters in the domain are formed. In the second stage, peripheral clusters are formed, while core clusters continue to grow. In the third stage, core clusters grow again. By focusing on the elementary process in clustering, we can understand the structure of citation networks in each research domain and judge the location of each cluster and paper, which cannot be seen in the final clustering results.

Emergence of Scaling in Random Networks

Article

Jan 1999

Emergence of Scaling in Random Networks

Article

Jan 1999

Systems as diverse as genetic networks or the World Wide Web are best described as networks with complex topology. A common property of many large networks is that the vertex connectivities follow a scale-free power-law distribution. This feature was found to be a consequence of two generic mech-anisms: (i) networks expand continuously by the addition of new vertices, and (ii) new vertices attach preferentially to sites that are already well connected. A model based on these two ingredients reproduces the observed stationary scale-free distributions, which indicates that the development of large networks is governed by robust self-organizing phenomena that go beyond the particulars of the individual systems.

The Matthew effect in science

Article