You are on page 1of 38

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/295253741

On the Shoulders of Giants

Chapter · January 2003


DOI: 10.1007/978-1-4471-0051-5_5

CITATION READS
1 3,582

1 author:

Chaomei Chen
College of Computing and Informatics, Drexel University
390 PUBLICATIONS   8,081 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Journal: Information Visualization View project

Research: A Visual Analytic Observatory of Scientific Knowledge (VAO) View project

All content following this page was uploaded by Chaomei Chen on 06 April 2016.

The user has requested enhancement of the downloaded file.


Chapter 5 The Structure and Dynamics of Scien-
tific Knowledge

If I have seen further it is by standing on the shoulders of Giants.


Isaac Newton (1642-1727)

In a letter to Robert Hooke in 1675, Isaac Newton made his most famous
statement: “If I have seen further it is by standing on the shoulders of Giants.”
This statement is now often quoted to symbolize scientific progress. Robert Mer-
ton examined the origin of this metaphor in his On the Shoulders of Giants
(Robert K. Merton, 1965). The shoulders-of-giants metaphor can be traced to the
French philosopher Bernard of Chartres, who said that we are like dwarfs on the
shoulders of giants, so that we can see more than they, and things at a greater dis-
tance, not by virtue of any sharpness of sight on our part, or any physical distinc-
tion, but because we are carried high and raised up by their giant size.
In a presentation at the Conference on The History and Heritage of Science In-
formation Systems at Pittsburgh in 1998, Eugene Garfield used “On the Shoulders
of Giants” as the title of his tributes to an array of people who had made tremen-
dous contributions to citation indexing and science mapping, including Robert
King Merton, Derek John de Solla Price (1922-1983), Manfred Kochen (1928-
1989), Henry Small, and many others (Eugene Garfield, 1998). In 1999, Henry
Small used On the Shoulders of Giants to entitle his ASIS Award Speech (Henry
Small, 1999). He explained that if a citation can be seen as standing on the shoul-
der of a giant, then co-citation is straddling the shoulders of two giants, a pyramid
of straddled giants is a specialty, and a pathway through science is playing leap-
frog from one giant to another. Henry Small particularly mentioned Belver Grif-
fith (1931-1999) and Derek Price as the giants who shared the vision of mapping
science with co-citation. Griffith introduced the idea of using multidimensional
scaling to create a spatial representation of documents. According to Small, the
work of Derek Price in modeling of the research front (Price, 1965) had a major
impact on his thinking.
The goal of this chapter is to introduce some landmark works of giants in quan-
titative studies of science, especially groundbreaking theories, techniques, and ap-
plications of science mapping. Henry Small praised highly the profound impact of

1
Thomas Kuhn on visualizing the entire body of scientific knowledge (Henry
Small, 1999). He suggested that if Kuhn's paradigms are snapshots of the structure
of science at specific points in time, examining a sequence of such snapshots
might reveal the growth of science. Kuhn (1970) speculated that citation linkage
might hold the key to solve the problem.
In this chapter, we start with general descriptions of science in action as re-
flected through indicators such as productivity and authority. We follow the de-
velopment of a number of key methods to science mapping over the last few dec-
ades, including co-word analysis and co-citation analysis. These theories and
methods have been an invaluable source of inspiration for generations of re-
searchers across a variety of disciplines. And we are standing on the shoulders of
giants.

5.1 Matthew Effect


What is the nature of scholarly publishing? Will the Internet-led electronic pub-
lishing fundamentally change it? Michael Koenig and Toni Harrell (Koenig &
Harrell, 1995) addressed this issue by using Derek Price’s urn model of Lotka’s
law.
In 1926, Alfred Lotka (1880–1949) found that the frequency distributions of
authors' productivity in chemistry and physics followed a straight line with a slope
of 2:1 (Lotka, 1926). In other words, the number of authors who published N pa-
pers is about twice the number of authors who published 2N papers. This is
known now as Lotka’s law.
Derek Price illustrated the nature of scholarship with the following urn model
(D. Price, 1976). To play the game, we need a bag, or an urn, and two types of
balls labeled "S" for success or "F" for failure. The player’s performance in the
game is expected to track the performance of a scholar. The scholar must publish
one paper to start the game. Whenever he draws an “F”, the game is over. There
are two balls at the beginning of the game: one “S” and the other “F”. The odd is
50-50 on the first draw. If he draws an "S", this ball plus another “S” ball will be
put in the bag and the scholar can make another draw. The odds improve with
each round of success. This game can replicate almost exactly the distribution that
Lotka derived from observation.
Price’s urn model accurately and vividly characterizes the nature of scholar-
ship. A scholar is indeed playing a game: Publications and citations are how
scholars score in the game (Koenig & Harrell, 1995). To stay in the game, schol-
ars must play it successfully. Each publication makes it easier for the scholar to
score again. Success breeds success Electronic publishing on the Internet has the
potential to increase the odds in the urn because it has the potential to speed up the
process. Can online accessibility boost the citations of an article? Steven Law-
rence and his colleagues found a strong correlation between the number of cita-
tions of an article and the likelihood that the article is online (Lawrence, 2001).
They analyzed 119,924 conference articles in computer science and related disci-

2
plines, obtained from DBLP (http://dblp.uni-trier.de), an online computer science
bibliography. Citation counts and online availability were estimated using Re-
searchIndex. Their conclusion was that online articles were likely to acquire more
citations.
Robert King Merton is an American sociologist who has revolutionized sociol-
ogy and mass communication. He is a pioneer in the sociology and the history of
sciences. He drew our attention to the “Matthew Effect” in scientific communities
(R. K. Merton, 1968). He adopted the term from St. Matthew’s Gospel in the Bi-
ble: “For unto everyone that hath shall be given, and he shall have abundance; but
for him that hath not shall be taken away even that which he hath.” ( Bible, Mat-
thew 13:12, 25:29). "Matthew Effect" sums up the phenomenon that the rich
get richer and the poor get poorer. In the context of science, “the richness” refers
to the reputation and prominence of an established scientist; in contrast, “the
poor” includes scientists who have not reached this level. Established scientists
tend to receive more than their fair share of credits at the expense of those who are
not famous. Here is how Merton described the Matthew effect in scientific reward
systems:
“You usually notice the name that you’re familiar with. Even if it’s last, it will
be the one that sticks. In some cases, all the names are unfamiliar to you, and
they’re virtually anonymous. But what you note is the acknowledgement at the end
of the paper to the senior person for his ‘advice and encouragement.’ So you will
say: ‘This came out of Greene’s lab, or se and-so’s lab.’ You remember that, ra-
ther than the long list of authors.”
Social and political forces may limit the recognition of a scientist. Merton de-
scribed the "41st chair” phenomenon in the French Academy, which can only al-
low a maximum of 40 members. Many talented individuals were denied a mem-
bership of the Academy simply because of this restriction.
Merton’s another contribution to sociology of science is the concept of scien-
tific obliteration. He first described the idea in On the Shoulders of Giants (Robert
K. Merton, 1965):
“Natural enough, most of us tend to attribute a striking idea or formulation to
the author who first introduced us to it. But often, that author has simply adopted
or revived a formulation which he (and others versed in the same tradition) knows
to have been created by another. The transmitters may be so familiar with its ori-
gins that they mistakenly assume these to be well known. Preferring not to insult
their readers' knowledgeability, they do not cite the original source or even refer
to it. And so it turns out that the altogether innocent transmitter becomes identi-
fied as the originator of the idea when his merit lies only in having kept it alive, or
in having brought it back to life after it had long lain dormant or perhaps in hav-
ing put it to new and instructive use.”
Obliteration happens in a scientific reward system when researchers no longer
feel necessary to cite something everyone has already taken for granted. Take Ar-
chimedes’ constant  for example. Archimedes discovered the ratio between the
diameter and circumference of a circle: . As Archimedes’ constant becomes in-

3
creasingly familiar even to schoolchildren, scientists would cite Archimedes’ pri-
mordial paper less and less, until finally there is no need to cite it at all, which
means his original paper would have been obliterated. This is regarded as one of
the highest compliments the community of scientists can pay to a scientist because
of a contribution that was so basic, so vital, and so well known that every scientist
can simply take it for granted (E. Garfield, 1975).
Just to mention two more examples of obliteration. One is the notion of “the
exponential growth of scientific literature”. Derek Price formulated the law of ex-
ponential growth of scientific literature in 1950 in his paper to the 6th International
Congress for the History of Science at Amsterdam. Before long, scientists from
different disciplines obliterated it and took the exponential growth for granted.
The notion of “paradigm shift” is another example. Phrases such as “new para-
digms” and “a paradigm shift” frequently appear in scientific literature without di-
rect citations to Thomas Kuhn’s seminal book The Structure of Scientific Revolu-
tion (Kuhn, 1962).
In information science, an “obliteration” hallmark is the annual Award of Mer-
its from the American Society for Information Science and Technology
(ASIS&T). The Award of Merit is the highest honor of ASIS&T for individuals
who have made an outstanding contribution to the field of information science.
Henry Small of ISI was the recipient of the 1999 award for his work in co-citation
analysis. We will include some examples of his work in this chapter. Don Swan-
son, professor emeritus at the University of Chicago, was the recipient of the 2000
for his renowned work in undiscovered public knowledge. The next chapter will
include a study of his research. Table 8.1 lists recipients of the award since 1964.
In Science since Babylon, Derek Price (1961) used the term invisible college to
emphasize the role of informal networks of scientists in scientific communication.
The term was originally used in the 17th century’s London to refer to an informal
club of artisans and practitioners before the formal organization of the Royal So-
ciety. Blaise Cronin (1972) regarded such informal scholarly communication net-
works as the "lifeblood of scientific progress for both the physical and the social
sciences.”
Science mapping has been a long-lasting pursuit for revealing the dynamics of
an invisible college and the evolution of intellectual structures. Derek Price has
been regarded as the leader in the field of Science of Science, which is a precursor
of the social studies of science and the field of scientometrics. Scientometrics is
the quantitative study of scientific communications.
In science mapping, we must consider a wide variety of fundamental concepts
that distinguish the level of granularity of each individual study. Such concepts
are known as units of analysis. Examples of abstract units include ideas, concepts,
themes, and paradigms. These concepts are represented and conveyed through
words, terms, documents, and collections by individual authors, groups of authors,
specialties, and scientific communities. The following examples in this chapter il-
lustrate association relationships between several types of units of analysis, such
as word co-occurrences in text, document co-occurrences in bibliography (docu-

4
ment co-citation), author co-occurrences in bibliography (author co-citation), and
patent occurrences in patent publications (patent co-citation).
Science mapping reveals structures hidden in scientific literature. The defini-
tion of association determines the nature of the structure to be extracted, to be vis-
ualized, and to be eventually interpreted. Co-word analysis (Michel Callon, Law,
& Rip, 1986) and co-citation analysis (Henry Small, 1973) are among the most
fundamental techniques for science mapping. Small (H. S. Small, 1988) described
the two as follows: "if co-word links are viewed as translations between problems,
co-citation links have been viewed as statements relating concepts." They are the
technical foundations of the contemporary quantitative studies of science. Each
offers a unique perspective on the structure of scientific frontiers. Researchers
have found that a combination of co-word and co-citation analysis could lead to a
clearer picture of the cognitive content of publications (Braam, Moed, & Raan,
1991a, 1991b).

5.2 Maps of Words


The tradition of deriving higher-level structures from word-occurrence patterns
in text originated in the co-word analysis method developed in the 1980’s (M.
Callon, Courtial, Turner, & Bauin, 1983; Michel Callon, et al., 1986). Co-word
analysis is a well-established camp in scientometrics, which is a field of quantita-
tive studies of science concerning with indicators and metrics of the dynamics of
science and technology at large. The outcome of co-word analysis was typically
depicted as a network of concepts.

5.2.1 Co-Word Maps


The history of co-word analysis has some interesting philosophical and socio-
logical implications for what we will see in later chapters. First, one of the key ar-
guments of the proponents of co-word analysis is that scientific knowledge is not
merely produced within “specialist communities” which independently define
their research problems and delimit clearly the cognitive and methodological re-
sources to be used in their solution. The attention given to “specialist communi-
ties” is due to the influence of the work done by Thomas Kuhn, particularly in his
Postscript to the second edition of The Structure of Scientific Revolutions. There
are some well-known examples of this approach, notably the invisible college by
Diana Crane (1972). The specialty areas are often identified by an analysis of cita-
tions in scientific literature (E. Garfield, Malin, & Small, 1978). Co-citation anal-
ysis has been developed in this context (H. Small, 1977; H. Small & Greenlee,
1980). A general criticism of the sociology of specialist communities was made
by Knorr-Cetina and Rip. Edge (1979) gave critical comments on delimiting spe-
cialty areas by citations. In 1981, the issue 11(1) of Social Studies of Science was
devoted to the analysis of scientific controversies. We will return to Kuhn’s theory
when we explain its roles in visualizing scientific frontiers in later chapters of the
book.

5
In 1976, Henry Small raised the question of social-cognitive structures in sci-
ence and underlined the difficulties of using experts to help identify them. This is
because experts are biased. Co-word analysis was developed to provide an “objec-
tive” approach without the help of domain experts.
The term leximappe was used to refer to this type of concept maps. More spe-
cific types of such maps are inclusion maps and proximity maps. Subsequent de-
velopments in relation to co-word analysis have incorporated artificial neural net-
work techniques such as self-organized maps to depict patterns and trends derived
from text. See (Lin, 1997; Noyons & van Raan, 1998) for example.
The pioneering software for concept mapping is Leximappe, developed in
1980s. It organizes a network of concepts based on associations determined by the
co-word method. In 1980's, it was Leximappe that had turned co-word analysis in-
to an instrumental tool for social scientists to carry out numerous studies originat-
ed from the famous the actor-network theory (ANT).
Key concepts in Leximappe include poles and their position in concept maps.
The position of the poles is determined by centrality and density. The centrality
implies the capacity of structuring; the density reflects the internal coherence of
the pole.
Leximappe is used to create structured graphic representations of concept net-
works. In such networks, vertices represent concepts; the strength of the connec-
tion between two vertices reflects the strength of their co-occurrence. In the early
days, an important step was to tag all words in the text as a noun, a verb, or an ad-
jective. Algorithms used in information visualization systems such as ThemeS-
cape (J. A. Wise, et al., 1995) have demonstrated some promising capabilities of
filtering out nouns from the source text.

5.2.2 Inclusion Index and Inclusion Maps


Inclusion maps and proximity maps are two types of concept maps resulted
from co-word analysis. Co-word analysis measures the degree of inclusion and
proximity between keywords in scientific documents and draws maps of scientific
areas automatically in inclusion maps and proximity maps, respectively.
Metrics for co-word analysis have been extensively studied. Given a corpus of
N documents, each document is indexed by a set of unique terms that can occur in
multiple documents. If two terms, ti and tj, appear together in a single document, it
counts as a co-occurrence. Let ck be the number of occurrences of term t k in the
corpus and cij be the number of co-occurrences of terms ti and tj, which is the
number of documents indexed by both terms. The inclusion index Iij is essentially
a conditional probability. Given the occurrence of one term, it measures the likeli-
hood of finding another term in documents of the corpus:
Iij = cij /min(ci , cj)
For example, Robert Stevenson’s Treasure Island has a total of 34 chapters.
Among them the word map occurred in 5 chapters, cmap = 5, and the word treasure
occurred 20 chapters, ctreasure = 20. The two terms co-occur in 4 chapters, thus cmap,

6
treasure=4. Imap, treasure
= 4/5=0.8. In this way, we can construct an inclusion matrix of
terms based on their co-occurrence. This matrix defines a network. An interesting
step described in the original version of co-word analysis is to remove certain
types of links from this network.
The original co-word analysis prunes a concept graph using a triangle inequali-
ty rule on conditional probabilities. Suppose we have a total of N words in the
analysis, for 1  i, j, k  N, ij, ik, and kj represent the weights of links in the
network and ij is defined as 1- Iij. Given a pre-defined small threshold , if there
exists an index k such that ij > ik*kj + , then we should remove the link Iij.
Because ik*kj defines the weight of a path from term ti to tj, what this operation
means is if we can find a shorter path from term ti to tj than the direct path, then
we choose the shorter one. In other words, if a link violates the triangle inequality,
it must be invalid; therefore, it should be removed. By rising or lowering the
threshold , we can decrease or increase the number of valid links in the network.
This algorithm is simple to implement. In co-word analysis, we usually only com-
pare a one-step path with a two-step path. However, when the size of the network
increases, this simple algorithm tends to allow in too many links and the resultant
co-word map tends to lose its clarity. In next chapter, we will introduce Pathfinder
network scaling as a generic form of the triangle inequality condition, which ena-
ble us compare much longer paths connecting two points and detect much subtle
association patterns in data.
Fig. 5.1 shows a co-word map based on the inclusion index. The co-word anal-
ysis was conducted on index terms of articles published in 1990 from a search in
the Web of Science with the query “mass extinction”. The meaning of this particu-
lar co-word map should become clear when you complete Chapter 7, which con-
tains a detailed account of the background and key issues in the study of mass ex-
tinction. The main reason we skip the explanation here is because of its
involvement with theories and examples of competing paradigms, a unique char-
acteristic of a scientific frontier.

7
Fig. 5.1. An inclusion map of research in mass extinction based on index terms of
articles on mass extinction published in 1990. The size of a node is proportional to
the total number of occurrences of the word. Links that violate first-order triangle
inequality are removed (=0.75).

5.2.3 The Ontogeny of RISC


Steve Steinberg (Steinberg, 1994) addressed several questions regarding the
use of a quantitative approach to identify paradigm shifts in the real world. He
chose to examine Reduced Instruction Set Computing (RISC). The idea behind
RISC was that a processor with only a minimal set of simple instructions could
outperform a processor that included instructions for complex high-level tasks. In
part, RISC marked a clear shift in computer architecture and had reached some
degree of consensus. Steinberg searched for quantitative techniques that could
help his investigation. Eventually he found the co-word analysis technique that
could produce a map of the field, a visualization of the mechanisms, and a battle
chart of the debate. He wrote (Steinberg, 1994): “If I could see the dynamics of a
technical debate, I thought, perhaps I could understand them.”
He collected all abstracts with the keyword RISC for the years 1980-1993 from
the INSPEC database, filtered out the 200 most common English words, and
ranked the remaining words by frequency. The top 300 most frequently occurred
words were given to three RISC experts to choose those words central to the field.
Finally, words chosen by the experts were aggregated by synonyms into 45 key-
word clusters. The inclusion index was used to construct a similarity matrix. This
matrix was mapped by MDS with ALSCAL. The font size of a keyword was pro-

8
portional to the word’s frequency. Strongly linked keywords were connected by
straight lines.
Fig. 5.2 shows the co-word map of the period of 1980-1985. The first papers to
explicitly examine and define RISC appeared within this period. The design phi-
losophy of RISC was so opposed to the traditional computing architecture para-
digm, every paper in this period was written to defend and justify RISC. The map
shows two main clusters. One is on the left, surrounding keywords such as regis-
ter, memory, simple, and pipeline. These are the architectural terms that uniquely
define RISC. The other cluster is on the right, centered on keywords such as lan-
guage and CISC. These are the words that identify the debate between the RISC
and CISC camps. Language is the most frequent keyword on the map. According
to Steinberg, the term language most clearly captures the key to the debate be-
tween RISC and CISC. While CISC proponents believed that a processor’s in-
struction set should closely correspond to high-level languages such as
FORTRAN and COBOL, RISC proponents argue that simple instructions were
better than high-level instructions. This debate is shown in the co-word map with
the connections between language, CISC, compiler, and programming.

Fig. 5.2. The co-word map of the period of 1980-1985 for the debate on RISC.
To illustrate the paradigm shift, we also include the co-word map of another
period: 1986-1987 (Fig. 5.3). During this period, Sun introduced the first com-
mercially important RISC microprocessor – the SPARC in 1986. RISC had trans-
formed from papers to a tangible product, backed by investors. The bi-polar co-
word map for the previous period is now predominated by the RISC cluster. The
technology of RISC implementation, namely VLSI, has become larger and more
central.

9
Fig. 5.3. The co-word map of another period: 1986-1987 for the debate on RISC.
On the one hand, the reconfiguration of the co-word map from bi-polar to lop-
sided indicates that the high-level language argument had been settled. On the
other hand, the map provides few clues of how this transformation took place. The
lack of interpretable indicators at detailed levels is not uncommon with co-word
maps and indeed with other types of bibliometric maps as well. In order to inter-
pret a visualized structure, one has to resort to some substantial levels of domain
knowledge, or at least one has to read some qualitative summaries of the subject.
In fact, it is advisable to consult a good review article to double-check the validity
of interpretations along the map. In this example, Steinberg himself was an expert
on the topic of RISC and he incorporated his domain knowledge into the interpre-
tation of co-word maps generated from abstracts. Researchers in quantitative stud-
ies of science have also recommended a multiple-approach strategy - to the same
phenomenon with a few different methods - so that one can compare and contrast
results from different perspectives and piece together a big picture. If mapping the
paradigm shift with one single technique is like the blind men approaching the el-
ephant, combining different techniques may lead to a more accurate model of the
elephant. Next, we turn to co-citation analysis, which is another major approach
that has been used for science mapping.

5.3 Co-Citation Analysis


Citation analysis takes into account one of the most crucial indicators of schol-
arship - citations. Citation analysis has a unique position in the history of science
mapping because several widely used analytical methods have been developed to
extract citation patterns from scientific literature and these citation patterns can
provide insightful knowledge of an invisible college. Traditionally, both philoso-
phy of science and sociology of knowledge have a strong impact on citation anal-
ysis. Opponents of citation analysis criticize its approach influenced by the idea of

10
invisible colleges and scientific communities, and argue that the way science op-
erates is far beyond the scope of citation practices (Michel Callon, et al., 1986).
However, this issue cannot be simply settled by theoretical arguments. Longitudi-
nal studies and large-scale domain analysis can provide insightful answers, but
they tend to be very time-consuming and resource demanding. In practice, re-
searchers have been exploring frameworks that can accommodate both co-word
analysis and co-citation analysis (Braam, et al., 1991a, 1991b). These efforts may
provide additional insights into the philosophical and sociological debates. Docu-
ment co-citation analysis (DCA) and author co-citation analysis (ACA) represent
the two most prolific mainstream approaches to co-citation analysis. Here we first
introduce DCA and then explain ACA.

5.3.1 Document Co-Citation Analysis


Citation indexing provides a device for researchers to track the history of ad-
vances in science and technology. One can trace a network of citations to find out
the history and evolution of a chain of articles on a particular topic. The goal of ci-
tation analysis is to make the structure of such a network more recognizable and
more accessible.
Traditional citation analysis is typically biased to journal publications due to
the convenience of available citation data. Expanding the sources to other scien-
tific inscriptions, such as books, proceedings, grant proposals, patents, preprints,
and digital resources on the Internet, has begun to attract the attention of research-
ers and practitioners. In 2002, when I wrote the first edition of the book, we antic-
ipated to see a sharp increase in patent analysis and studies utilizing Web-based
citation indexing techniques in the next 3~5 years because of the growing interest
and commercial investments in supporting patent analysis with knowledge dis-
covery and visualization techniques. Today, major resources for citation analysis
include Thomson Reuters’ Web of Science, Elsevier’s Scopus, and Google Schol-
ar. Fig. 5.4 shows a visualization of a document co-citation network of publica-
tions in Data and Knowledge Engineering. The color coded clusters indicate the
focus of the community at various stages of the research. The most cited paper is
the one that invented the entity-relationship modeling method by Peter Chen.

11
Fig. 5.4. A document co-citation network of publications in Data and Knowledge
Engineering.

5.3.1.1 Specialties
In information science, the term specialty refers to the perceived grouping of
scientists who are specialized on the same or closely related topics of research.
Theories of how specialties evolve and change started to emerge in the 1970s (H.
G. Small & Griffith, 1974). Researchers began to focus on the structure of scien-
tific literatures in order to identify and visualize specialties, although they did not
use the term “visualization” at that time.
Recently, science-mapping techniques have begun to reveal structures of scien-
tific fields in several promising visualization metaphors, including networks, land-
scapes, and galaxies. The ability to trace scientific and technological break-
throughs from these science maps is particularly important. The key questions are:
what are these maps telling us, and how do we make use of such maps at both
strategic and tactical levels?
Today’s most widely used citation index databases such as SCI and SSCI were
conceived in the 1950s, especially in Garfield’s pioneering paper published in Sci-
ence (E. Garfield, 1955). In the 1960s, several pioneering science mapping studies
began to emerge. For example, Garfield, Sher, and Torpie created the historical
map of research in DNA (E Garfield, H., & Torpie, 1964). Sher and Garfield
demonstrated the power of citation analysis in their study of Nobel Prize winners'
citation profiles (Sher & Garfield, 1966). Fig. 5.5 shows how citation analysis
spotted a vital missing citation to earlier work (E. Garfield, 1996).

12
Fig. 5.5. Citation analysis detected a vital missing citation from Mazur’s paper in
1962 to Rydon’s paper in 1952.
In the 1970s, information scientists began to focus on ways that can reveal pat-
terns and trends reflected through scientific literature. Henry Small demonstrated
the power of SCI-Map in mapping the structure of research in AIDS (H. Small,
1994). Once the user specified an author, a paper, or a key word as the seed, SCI-
Map could create a map of related papers by adding strongly co-cited papers to the
map. The creation of a map involved a series of iterations of clustering. The layout
was generated by a method called geometric triangulation, which is different from
the MDS approach used in Small’s earlier work and in similar studies in the US
and Europe (E. Garfield, 1996).
Henry Small and Belver Griffith initiated co-citation analysis for identifying
and mapping specialties from the structure of scientific literature (H. G. Small &
Griffith, 1974). Articles A and B have a co-citation count of k if there are k arti-
cles and each of them cites both articles A and B. The co-citation rate of A and B
is defined as the number of such instances. A high co-citation rate implies a strong
intellectual tie between two articles.
In a longitudinal study of collagen research, Henry Small tracked the move-
ment of specialties in collagen research using a cluster-based approach (H. Small,

13
1977). He emphasized the fundamental role of systematic and consistent methodo-
logical frameworks. He used the frequency of co-citation to measure the strength
of the association between articles on the topic. He marked clusters of highly cited
articles in MDS maps with contour lines so that he could track rapid sifts in re-
search focus from one year to another as articles moved in and out key cluster
contours and used it as an indicator of “revolutionary” changes.
In the 1980s, the Institute for Science Information (ISI) published the Atlas of
Science in Biochemistry and Molecular Biology, which identified more than 100
distinct clusters of articles, known as research front specialties, and provided a
distinct snapshot of scientific networks. The Atlas was constructed based on co-
citation relationships between publications in the field over a period of one year.
In 1989, Garfield and Small explained how software like SCI-Map could help us-
ers navigate the scientific literature and visualize the changing frontiers of science
based on citation relationships (E. Garfield & Small, 1989). Henry Small de-
scribed in detail his citation mapping approach to visualizing science (H. Small,
1999). Fig. 5.6 shows a global map of science for 1996 produced by co-citation
mapping. The map highlighted major connections among disciplines such as eco-
nomics, neuroscience, biomedicine, chemistry, and physics. The size of a circle
was made proportional to the volume of a particular scientific literature, for ex-
ample, the large biomedical circle in the center of the map indicating the huge
number of biomedicine publications in journals. Computer science, shown as a
relatively small circle in the map, linked to imaging and economics. The small
volume of computer science reflected the fact that journal publications are merely
a small proportion of the entire computer science literatures, typically including
conference proceedings, technical reports, and preprints. One can also zoom into
the global map of science and examine local structures (See Fig. 5.7 and Fig. 5.8).

14
Fig. 5.6. A global map of science based on document co-citation patterns in 1996,
showing a linked structure of nested clusters of documents in various disciplines
and research areas. Reproduced form (Garfield, 1998).

Fig. 5.7. Zooming in to reveal a detailed structure of biomedicine (Garfield,


1998).

15
Fig. 5.8. Zooming in even further to examine the structure of immunology (Gar-
field, 1998).
MDS maps and clustering algorithms are typically used in co-citation analysis
to represent co-citation structures. There is an increasing interest in using graph-
drawing techniques to depict the results of co-citation analysis, including mini-
mum spanning trees (MST) and Pathfinder networks (PF). The increased use of
the metaphor of an information landscape is another trend, in which the entire
structure can be rendered as a mountain terrain or a relief map.
5.3.1.2 Narratives of Specialties
Creating a science map is the first step towards exploring and understanding
scientific frontiers. Science maps should guide us from one topic or specialty to
related topics or specialties. Once we have a global map in our hands, the next
logical step is to find out how we can make a journey from one place to another
based on the information provided by the map. Small introduced the concept of
passage through science (H. Small, 1999). Passages are chains of articles in scien-
tific literature. Chains running across the literature of different disciplines are like-
ly to carry a method established in one discipline into another. Such chains are
vehicles for cross-disciplinary fertilization. Traditionally, a cross-disciplinary
journey would require scientists to make a variety of connections, translations,
and adaptations. Small demonstrated his powerful algorithms by blazing a magnif-
icent trail of more than 300 articles across the literatures of different scientific dis-
ciplines. This trailblazing mechanism development has brought Bush’s (Bush,
1945) concept of information trailblazing to life.
Henry Small described what he called the synthesis of specialty narratives from
co-citation clusters (H. Small, 1986). This paper won the JASIS best-paper award
in 1986. Small first chose a citation frequency threshold to select the most cited
documents in SCI. The second step was to determine the frequency of co-citation
between all pairs of cited documents above the threshold. Co-citation counts were
normalized by Salton’s cosine formula. Documents were clustered using the sin-
gle-link clustering method, which was believed to be more suitable than the com-

16
plete-link clustering algorithm because the number of co-citation links can be as
many as tens of thousands. Single-link clusters tend to form a mixture of densely
and weakly linked regions in contrast to more densely packed and narrowly fo-
cused complete-link clusters. MDS was used to configure the layout of a global
map. Further, Small investigated how to blaze trails in the knowledge space repre-
sented by the global map. He called this type of trail the specialty narrative.
Small addressed how to transform a co-citation network into a flow of ideas.
The goal for specialty narrative construction is to find a path through such net-
works so as to track the trajectory of scientists who had encountered these ideas.
Recall that the traveling salesman problem (TSP) requires the salesman to visit
each city exactly once along a route optimized against a given criterion. We are in
a similar situation with the specialty narrative construction, or more precisely, the
re-construction of narrative trails when we retrace the possible sequence of
thought by following trails of co-citation links. TSP is a hard problem to solve.
Luckily, there are some very efficient algorithms to traverse a network, namely
breath-first search (BFS) and depth-first search (DFS). Both result in a minimum-
spanning tree (MST). Small considered several possible heuristics for the traversal
in his study. For example, when we survey the literature, we tend to start with
some old articles so as to form a historical context. A reasonable approach is to
start from the oldest article in the co-citation network. In this example, DFS was
used to generate an MST. The longest path through the MST tree was chosen as
the main sequence of the specialty narrative (See Fig. 5.9).

Fig. 5.9. The specialty narrative of leukemia viruses. Specialty narrative links are
labeled by citation-context categories. Reproduced from (H. Small, 1986).

17
The context of citing provides first-hand information on the nature of citation.
A specialty narrative is only meaningful and tangible if sufficient contextual in-
formation of citation is attached to the narrative. The citation context of a given
article consists of sentences that explicitly cite the article. Such sentences may
come from different citing articles. Different authors may cite the same article for
different reasons. On the other hand, researchers may cite several articles within
one sentence. Small took all these circumstances into account in his study. In the
foreseeable future, we will still have to rely on human intervention to make such
selection, as opposed to automated algorithmic devices. Nevertheless, NEC’s Re-
searchIndex has shown some promising signs of how much we might benefit from
citation contexts automatically extracted from documents on the Web. In his 1986
specialty narrative study, Small had to examine passages from citing papers, cod-
ed them, and keyed them before running a program to compute the occurrence
frequencies. This specialty narrative was rigorously planned, carefully carried out,
and thoroughly explained. Henry Small’s JASIS award-wining paper has many in-
spiring ideas and technical solutions that predated the boom of information visual-
ization in the 1990s. Over the last 15 years, this paper has been a source of inspi-
ration for citation analysis; we expect it will also influence information
visualization and knowledge visualization in a fundamental way.
Robert Braam, Henk Moed and Anthony van Raan investigated whether co-
citation analysis indeed provided a useful tool for mapping subject-matter special-
ties of scientific research (Braam, et al., 1991a, 1991b). Most interestingly, the
cross-examination method they used was co-word analysis. Their work clarified a
number of issues concerning co-citation analysis.
The cluster of co-cited documents is considered to represent the knowledge
base of a specialty (H. Small, 1977). In a review of bibliometric indicators, Jean
King (King, 1987) sums up objections against co-citation analysis: loss of relevant
papers, inclusion of non-relevant papers, overrepresentation of theoretical papers,
time lag, and subjectivity in threshold setting. There were more skeptical claims
that co-citation clusters were mainly artifacts of the applied technique having no
further identifiable significance. Braam and his co-workers addressed several is-
sues in their investigation in response to such concerns. For example, does a co-
citation cluster identify a specialty? They used concepts such as “cognitive coher-
ence” within clusters and “cognitive differences” between clusters. Their results
suggested that co-citation analysis indeed showed research specialties, although
one specialty may be fragmented across several different clusters. They concluded
that co-citation clusters were certainly not artifacts of an applied technique. On the
other hand, their study suggested that co-citation clusters did not represent the en-
tire body of publications that comprised a specialty. Therefore, they concurred the
recommendation of Mullins et al. (Mullins, Snizek, & Oehler, 1988) that it would
be necessary to analyze different structural aspects of publications so as to gener-
ate significant results in science mapping.

18
5.3.2 Author Co-Citation Analysis
The 1980s saw the beginning of what turned out to be a second fruitful line of
development in the use of citation to map science - author co-citation analysis
(ACA). Howard White and Belver Griffith introduced ACA in 1981 as a way to
map intellectual structures (H. D. White & Griffith, 1981). The unit of analysis in
ACA is authors and their intellectual relationships as reflected through scientific
literatures. The author-centered perspective of ACA led to a new approach to the
discovery of knowledge structures in parallel to approaches used by document-
centered co-citation analysis (DCA).
5.3.2.1 Intellectual Structures
An author co-citation network offers a useful alternative starting point for co-
citation analysis, especially when we encounter a complex document co-citation
network, and vice versa. Katherine McCain (1990) gave a comprehensive tech-
nical review of mapping authors in intellectual spaces. ACA reached a significant
turning point in 1998 when White and McCain (Howard D. White & McCain,
1998b) applied ACA to information science in their thorough study of the field.
Since then ACA has flourished and has been adopted by researchers across a
number of disciplines beyond the field of citation analysis itself. Their paper won
the best JASIS paper award. With both ACA and DCA at our hands, we begin to
find ourselves in a position to compare and contrast messages conveyed through
different co-citation networks of the same topic as if we were having two pairs of
glasses.
Typically, the first step is to identify the scope and the focus of ACA. The raw
data are either analyzed directly or, more commonly, converted into a correlation
matrix of co-citation. Presentations often combine MDS with cluster analysis or
PCA. Groupings are often produced by hierarchical cluster analysis. Fig. 5.10 il-
lustrates a generic procedure of a standard co-citation procedure. For example,
node placement can be done with MDS; clustering can be done with the single- or
complete-link clustering; PCA might replace clustering. In practice, some re-
searchers choose to work on raw co-citation data directly, whereas others prefer to
work on correlation matrices. To our knowledge, there is no direct comparison be-
tween the two routes in terms of the quality of clustering, although it would be
useful to know the strengths and weaknesses of each route. Partition can divide a
global view into more manageable regions and make the map easier to understand.
Finally, additional information such as citation counts and co-citation strengths
can be rendered in the map to convey the message clearly.

19
Fig. 5.10. A generic procedure of co-citation analysis. Dashed lines indicate visu-
alization options.
In their pioneering 1981 study, White and Griffith created the first ever author
co-citation map of information science from the Social Sciences Citation Index 
(SSCI) for 1972-1979. Their map showed five main clusters of author within the
field of information science (See Fig. 5.11). Each cluster corresponded to a spe-
cialty:
1. Scientific communication,
2. Bibliometrics,
3. Generalists,
4. Document Analysis / Retrieval evaluation / Systems, and
5. Precursors.
In this first author co-citation map of information science, scientific communi-
cation is on the left and information retrieval on the right. Over the last twenty
years, researchers have created several co-citation maps of the field of information
science – their home discipline. Later maps have shared some characteristics of
this structure.

20
Fig. 5.11. The first map of author co-citation analysis, featuring specialties in in-
formation science (1972-1979) (White & Griffith, 1981).
The author co-citation map produced by White and Griffith (1981) depicted in-
formation science over a 5-year span (1972-1979). In 1998, 17 years later, White
and McCain (1998) generated a new map of information science based on a con-
siderably expanded 23-year span (1972-1995). They first selected authors who
had been highly cited in 12 key journals of information science. Co-citations of
120 selected authors between 1972 and 1995 were extracted from SSCI. They
generated maps of the top 100 authors in the field. Major specialties in the fields
were identified using factor analysis. The resultant map showed that the field of
information science consisted of two major specialties with little overlap in terms
of their memberships, namely experimental retrieval and scientific communica-
tion. Citation analysis belongs to the same camp as scientific communication. One
of remarkable findings was that the new map preserved some of the basic struc-
ture from the 1981 map: scientific communication on the right and information re-
trieval on the left.
White and McCain demonstrated that authors might simultaneously belong to
several specialties. Instead of clustering authors into mutual exclusive specialties,
they used PCA to accommodate the multiple-specialty membership for each au-
thor. First, the raw co-citation counts were transformed into Pearson's correlation
coefficients as a measure of similarity between pairs of authors (Howard D. White
& McCain, 1998b). They generated an MDS-based author co-citation map of 100
authors in information science for the period of 1972-1995. It is clear from the
map that information science was made of two major camps: the experimental re-
trieval camp on the right and the citation analysis camp on the left. The experi-
mental retrieval camp includes names such as Vannevar Bush (1890-1974), Ger-

21
ald Salton (1964-1988), and Don Swanson, whereas the citation camp includes
David Price (1922-1983), Eugene Garfield, Henry Small, and Howard White.
Thomas Kuhn (1922-1996) appears at about the coordinates of (–1.3, -0.8).
Since 1997, I started to explore Pathfinder network scaling as a vehicle to visu-
alize complex networks (C. Chen, 1997b, 1998b). See Figures 5.12-5.14. Path-
finder network scaling filters out excessive links in a network while maintaining
the salient structure of the network, more precisely, by preserving links that satisfy
the triangular inequality throughout the network. In 1999, I published a study of
author co-citation networks using Pathfinder network scaling techniques and
demonstrated the advantages of using Pathfinder over multidimensional scaling
because Pathfinder networks display connections explicitly and preserve salient
structure while pruning excessive links (C. Chen, 1999b).

Fig. 5.12. A two-dimensional Pathfinder network integrated with information on


term frequencies as the third dimension. Source: (C. Chen, 1998a).

22
Fig. 5.13. A Pathfinder network of SIGCHI papers based on their content similari-
ty. The interactive interface allows users to view the abstract of a paper seamlessly
as they navigate through the network. Source: (C. Chen, 1998a).

Fig. 5.14. A Pathfinder network of co-cited authors of the ACM Hypertext confer-
ence series (1989-1998). Source: (Chaomei Chen & Les Carr, 1999).
In 2003, Howard White revisited the same dataset used in their 1998 author co-
citation analysis and applied the Pathfinder network techniques to represent co-
cited authors. He concluded that Pathfinder networks provide considerable ad-
vantages over MDS maps because Pathfinder networks make the connections ex-
plicit. Fig. 5.15 shows a Pathfinder of 121 information science authors based on
raw co-citation counts. Garfield, Lancaster, and Salton are the most prominent au-

23
thors in the Pathfinder network; each is surrounded by a large number of co-cited
authors.

Fig. 5.15. A Pathfinder network of 121 information science authors based on raw
co-citation counts. Source: (Howard D. White, 2003).
White and McCain (Howard D. White & McCain, 1998a) discussed some is-
sues concerning detecting paradigm shifts. They compared author co-citation net-
works over three consecutive periods using INDSCAL. White and McCain's work
is a significant step towards understanding how we may grasp the dynamic of a
scientific community and track the development of a discipline.
5.3.2.2 Generalized Similarity Analysis
Generalized Similarity Analysis (GSA) is a generic framework for structuring
and visualizing distributed hypermedia resources (C. Chen, 1997a, 1998b). See
Chapter 4 for a detailed discussion. GSA uses Pathfinder networks to achieve an
improved clarity of a generic network. John Leggett of Texas A&M was a keynote
speaker at the 8th ACM Hypertext conference in Pittsburgh and he was talking
about “camps” in hypertext research and “runners” between these invisible camps:
who they are and where they are now. Inspired by White and McCain’s author co-
citation maps and John Leggett’s thought-provoking keynote speech, we were
able to pull things together by applying GSA to ACA. Leslie Carr at the Universi-
ty of Southampton provided me with the citation data for the ACM Hypertext con-
ference series. We presented a Pathfinder-powered visualization of the co-citation
networks of hypertext research at the 9th ACM Hypertext conference at Darmstadt
in Germany in 1999. Since then, we have developed a systematic and consistent
framework for ACA and document co-citation analysis (DCA) to accommodate

24
Pathfinder networks side-by-side with traditional dimensionality reduction tech-
niques such as MDS and PCA, and working with information visualization tech-
niques such as animation, color mapping, and three-dimensional landscaping. By
2001, we consolidated the methodology into a four-step procedure for domain
visualization (C. Chen & Paul, 2001). Having created global thematic landscapes
of a subject domain, our focus turned to the question of the functionality of such
visualizations and maps. It became clear that a more focused perspective is the
key to a more fruitful use of such visualizations. This is the reason we will turn to
Thomas Kuhn’s puzzle-solving paradigms and focus on the scenarios of compet-
ing paradigms in scientific frontiers in next chapter. Henry Small’s specialty nar-
rative also provides an excellent example of how domain visualization can guide
us to a greater access to the core knowledge in scientific frontiers.
5.3.2.3 MDS, MST, and Pathfinder
Multidimensional scaling (MDS) maps are among the most widely used ones to
depict intellectual groupings. MDS-based maps are consistent with Gestalt princi-
ples - our perceived groupings are largely determined by proximity, similarity,
and continuity. MDS is designed to optimize the match between pairwise proximi-
ty and distance in high-dimensional space. In principle, MDS should place similar
objects next to each other in a two- or three-dimensional map and keep dissimilar
ones farther apart.
MDS is easily accessible in most statistical packages such as SPSS, SAS, and
Matlab. However, MDS provides no explicit grouping information. We have to
judge proximity patterns carefully in order to identify the underlying structure.
Proximity-based pattern recognition is not easy and sometimes can be misleading.
For example, one-dimensional MDS may not necessarily preserve a linear rela-
tionship. A two-dimensional MDS configuration may not be consistent with the
results of hierarchical clustering algorithms - two points next to each other in an
MDS configuration may belong to different clusters. Finally, three-dimensional
MDS may become so visually complex that it is hard to make sense of it without
rotating the model in a 3D space and studying it from different angles. Because of
these limitations, researchers often choose to superimpose additional information
over an MDS configuration so as to clarify groupings of data points, for example,
by drawing explicit boundaries of point clusters in an MDS map. Most weakness-
es of MDS boil down to the lack of local details. If we treat an MDS as a graph,
we can easily compare the number of links across various network solutions and
an MDS configuration (See Table 5.1).
Table 5.1. Comparisons of networks by the number of links, where K is the num-
ber of unique edges in the graph.
G=(Vertice #Ver- #Eedges Ex-
s, Edges) tices ample:
N=367
MDS N 0 0
MST N N-1 366

25
G=(Vertice #Ver- #Eedges Ex-
s, Edges) tices ample:
N=367
PF N 3N 398
Full Matrix N  N(N-1)/2 61,17
5
Fig. 5.16 shows a minimum spanning tree (MST) of an author co-citation net-
work of 367 prominent authors in the field of hypertext. The original author co-
citation network consisted of 61,175 links among these authors. A fully connected
symmetric matrix of this size would have a maximum of 66,978 links, excluding
self-citations. In other words, the co-citation patterns were about 91% of the max-
imum possible connectivity. The MST solution selected a total of 366 strongest
links. It produces a much-simplified picture of the patterns.

Fig. 5.16. A minimum spanning tree solution of the author co-citation network
based on the ACM Hypertext dataset (Nodes=367, Links=366).
MST provides explicit links to display a more detailed picture of the underly-
ing network. If the network contains equally weighted edges, one can arbitrarily
choose any one of the MSTs. However, an arbitrarily chosen MST destroys the
semantic integrity of the original network because the selection of an MST is not
based on semantic judgments. Pathfinder network scaling resolves this problem by
preserving the semantic integrity of the original network. When geodesic distanc-
es are used, a Pathfinder network is the set union of all possible MSTs. Pathfinder
selects links by ensuring that selected links do not violate the triangle inequality
condition.

26
Fig. 5.17 is a Pathfinder network solution of the same author co-citation matrix.
Red circles mark the extra links when comparing to an MST solution. A total of
398 links were included in the network – the pathfinder network was 32 links
more than the number of links in its MST counterpart solution. These extra links
would be denied in MST because they form cyclic paths, but forming a cyclic path
alone as a link selection criterion may overlook potentially important links.

Fig. 5.17. The author co-citation network of the ACM Hypertext data in a Path-
finder network (Nodes= 367, Links=398).
In order to incorporate multiple aspects of author co-citation networks, we em-
phasize the significance of the following aspects of ACA (see Fig. 5.18):
 Represent an author co-citation network as a Pathfinder network;
 Determine specialty memberships directly from the co-citation matrix us-
ing PCA;
 Depict citation counts as segmented bars, corresponding to citation
counts over several consecutive years.

27
Fig. 5.18. The procedure of co-citation analysis as described in (Chen & Paul,
2001).
The results from the three sources, namely, Pathfinder network scaling, PCA,
and annual citation counts, are triangulated to provide the maximum clarity. Fig.
5.19 shows an author co-citation map produced by this method. This is an author
co-citation map of 367 authors in hypertext (1989-1998). PCA identified 39 fac-
tors, which corresponded to 39 specialties in the field of hypertext. Authors were
colored by factor loadings of the top three largest specialties. The strongest spe-
cialty was colored in red. The next two strongest ones were in green and blue, re-
spectively. The strongest specialty branches out from the top of the ring structure,
whereas the second strongest specialty appears to concentrate around the lower
left-hand corner of the ring.

Fig. 5.19. A Pathfinder network showing an author co-citation structure of 367


authors in hypertext research (1989-1998). The color of a node indicates its spe-

28
cialty membership identified by PCA: red for the most predominant specialty,
green the second, and blue the third. (© 1999 IEEE)
The colored PCA overlay allows us to compare structural positions of authors
and their presence in the three major specialties. Partitioning the network by color
provides a unique and informative alternative to traditional non-overlapping parti-
tions based on clustering and other mutually exclusive partition schemes. Less re-
stricted partition schemes are most appropriate when we deal with invisible col-
leges – identifying the membership of a scientist is rarely a clear cut. After all,
giants in scientific frontiers may well appear simultaneously in several specialties.
Fig. 5.20 shows a landscape view of the same author co-citation network en-
hanced by corresponding citation history of each author. Most cited authors be-
came landmarks in the scene. The shape of the invisible college associated with
this field of study began to emerge.

Fig. 5.20. A landscape view of the hypertext author co-citation network (1989-1998).
The height of each vertical bar represents periodical citation index for each author. (© 1999
IEEE).
We explored two types of animations: animations that display the distributions
of specialties and animations that display the growth of citation bars in the land-
scape. We chose to keep the underlying co-citation network constant, which
serves as a base map, and let the citation profiles grow. In effect, we have a grow-
ing thematic overlay within a static reference framework.
Applying Pathfinder network scaling to co-citation networks not only enriched
the applications of Pathfinder networks, but also let to deeper insights into the na-
ture of Pathfinder network scaling and how to interpret various patterns emerged
from such representations. Now we can systematically explain the meaning of a
co-citation network. For example, documents or authors in the centre or a relative-
ly fully connected area tend to be more generic and generally applicable, whereas

29
those located in peripheral areas of the Pathfinder network tend to represent more
specific topics.

5.4 HistCite
HistCite is a widely known example of historiography advocated by Eugene
Garfield for decades. HistCite is designed to depict citation connections between
scientific articles over time. It takes bibliographic records from the Web of Sci-
ence and generates a variety of tables and historiographic diagrams. In HistCite,
the number of citations of a reference in the entire Web of Science is called Glob-
al Citation Score (GCS), whereas the number of citations of a reference made by a
given set of bibliographic records, also known as a collection, is called Local Cita-
tion Score (LCS). Garfield has maintained a series of analyses using HistCite on
the web1.
In a historiography, published articles are organized according to the time of
their publication. Articles published in the earliest years are placed on the top of
the diagram, whereas more recent articles appear lower in the diagram. If article A
cites article B, then the connection is depicted by a directed line from A to B.
HistCite depicts how often an article has been cited by making the size of the node
proportional to the number of citations.
The historiograph in Fig. 5.21 illustrates what information could be conveyed
by such diagrams. The diagram is a co-citation historiograph generated by Eugene
Garfield on HistCite’s website2. According to Garfield, the dataset, or collection,
consists of papers that either have the words cocitation or co-citation in the titles
or cite one of three articles identified below:
 Small H., 1973, JASIS, 24(4), 265-269.
 Small H., 1974, Science Studies, 4(1), 17-40.
 Griffith BC, 1974, Science Studies, 4(4), 339-365.
To make it easier to read, I manually annotated the diagram by labeling nodes
with the lead author and a short phrase. The diagram makes it clear that the co-
citation research was pioneered by Henry Small in an article published in the
Journal of the American Society for Information Science (JASIS) in 1973. His ar-
ticle cited 20 articles, including Garfield’s 1955 paper in Science, which laid
down the foundation of citation analysis, Kessler’s 1963 paper, in which the con-
cept of bibliographic coupling was introduced.
In 1974, Small and Griffith further consolidated the conceptual foundation of
co-citation analysis. Garfield’s article on citation classics also appeared in the
same year. Three years later, Small deepened the co-citation methodology further
with a longitudinal cocitation study of collagen research. In the meantime, Mora-
vcsik and Murugesan studied function and quality of citations. Gilbert examined

1
http://garfield.library.upenn.edu/histcomp/
2
http://garfield.library.upenn.edu/histcomp/cocitation_small-
griffith/graph/2.html

30
the role of citations in persuasion. Small crystalized the notion of cited references
as concept symbols. The first major review article appeared in Library Trends in
1981, written by Linda Smith.
Author cocitation analysis (ACA) was first introduced in 1981 by Howard
White and Belver Griffith. Generally speaking, a co-citation network of authors
tends to be denser than a co-citation network of references, or document cocitation
analysis (DCA), as proposed by Small. ACA can be seen as an aggregated form of
cocitation analysis because different articles by the same author would be aggre-
gated to the name of the author in ACA. On the one hand, such aggregation may
simplify the overall complexity of the structure of a specialty. On the other hand,
such aggregation may also lose the necessary information to differentiate the
works by the same author. Considering it is quite common for a scholar to change
his or her research interests from time to time, one may argue that it would be
more informative to keep the distinct work of the same author separated instead of
lumping them altogether. The co-word analysis method was introduced by Callon
et al. in 1983. In 1985, Brooks investigated what motivated citers, while Small
further advanced the method for cocitation analysis with specific focus on the role
of cocitations as a clustering mechanism.
In 1987, Swanson’s work merged, leading to research in literature-based dis-
covery. His 1987 paper was followed by two more papers in 1990 and 1997, re-
spectively. In the meantime, Howard White and Katherine McCain reviewed the
state of the art of biometrics with their special focus on authors as opposed to oth-
er units of analysis. In 1998, White and McCain presented a comprehensive ACA
of information science and mapped the results in multidimensional scaling
(MDS). In 1999, we introduced Pathfinder network scaling to the analysis of au-
thor cocitation networks. In 2006, the publication of CiteSpace II marked a
streamlined analytic platform for cocitation studies. In 2008, the most recent addi-
tion to the landscape of cocitation literature was Martin Rosvall’s work that mod-
els information flows in networks in terms of random walks.

31
Fig. 5.21. An annotated historiograph of co-citation research. The original dia-
gram is at: http://garfield.library.upenn.edu/histcomp/cocitation_small-
griffith/graph/2.html

5.5 Patent Co-Citations


Patent analysis has a long history in information science, but recently there is a
surge of interest from the commercial sector. Numerous newly formed companies
are specifically aiming at the patent analysis market. Apart from historical driving
forces such as monitoring knowledge and technology transfer and staying in com-
petition, the rising commercial interest in patent analysis is partly due to the public
accessible patent databases, notably the huge amount of patent applications and
grants from the United States Patent and Trademark Office (USPTO). The public
can search patents and trademarks at USPTO’s website http://www.uspto.gov/ and
download bibliographic data from ftp://ftp.uspto.gov/pub/patdata/. Fig. 5.22
shows a visualization of a network of 1,726 co-cited patents.

32
Fig. 5.22. A minimum spanning tree of a network of 1,726 co-cited patents re-
lated to cancer research.
The availability of the abundant patent data, the increasingly widespread
awareness of information visualization, and the maturity of search engines on the
Web are among the most influential factors behind the emerging trend of patent
analysis. Many patent search interfaces allow users to search by specific sections
in patent databases, for example by claims. Statistical analysis and intuitive visu-
alization functions are by far the most commonly seen selling points from a
salesman’s patent analysis portfolio. The term visualization becomes so fashiona-
ble now in the patent analysis industry that from time to time we come across vis-
ualization software tools that turn out to be little more than standard displays of
statistics.
A particularly interesting example is from Sandia National Laboratory. Kevin
Boyack and his colleagues (2000) used their landscape-like visualization tool
VxInsight to analyze the patent bibliographic files from USPTO in order to an-
swer a number of questions. For example, where are competitors placing their ef-
forts? Who is citing our patents, and what types of things have they developed?
Are there emerging competitors or collaborators working in related areas? The
analysis was based on 15,782 patents retrieved from a specific primary classifica-
tion class from the US Patent database. The primary classification class is class
360 on Dynamic Magnetic Information Storage or Retrieval. A similarity measure
was calculated using the direct and co-citation link types of Small (1997). Direct

33
citations were given a weighting five times that of each co-citation link. These pa-
tents were clustered and displayed in a landscape view (See Fig. 5.23 and Fig.
5.24).

Fig. 5.23. Landscapes of patent class 360 for four five-year periods. Olympus’s patents
are shown in blue; Sony in green; Hitachi in green; Philips in magenta; IBM in cyan; and
Seagate in red. Reproduced from Figure 1 of (Boyack et al. 2000).

34
Fig. 5.24. Map of all patents issued by the US Patent Office in January 2000. De-
sign patents are shown in magenta; patents granted to universities in green; and
IBM’s patents in red. Reproduced from Figure 5 of (Boyack et al. 2000).

5.6 Summary
In this chapter, we have introduced factors that influence perceived impact of
scientific works, such as Mathew Effect. We focused on two mainstream ap-
proaches to science mapping, namely co-word analysis and co-citation analysis.
Within co-citation analysis, we distinguished document co-citation analysis and
author co-citation analysis. Key techniques used in and developed along with
these approaches were described, although our focus was on the fundamental re-
quirements and strategies rather than detailed implementations. More fundamental
issues were identified, that is, where should we go next from the global map of a
field of study from 60,000 feet above the ground? The central theme of this chap-
ter is on the shoulders of giants, which implies that the knowledge of the structure
of scientific frontiers in the immediate past holds the key to a fruitful exploration
of human being’s intellectual assets. Henry Small’s specialty narrative provided
an excellent example to mark the transition from admiring a global map to a more
detailed knowledge acquisition process. We conclude this chapter with a visuali-
zation of the literature of co-citation analysis. The visualization in Fig. 5.25 shows
a network of co-cited references from articles that cited either Henry Small or
Belver Griffith, the two pioneers of the co-citation research. The visualization is
generated by CiteSpace based on citations made between 1973 and 2011. The age
of an area of concentration is indicated by the colors of co-citation links. The ear-
lier works are in colder colors, i.e. blue. The more recent works are in warmer
colors, i.e. in orange. The upper half of the network was formed the first, whereas
the lower left area was the youngest. The network is divided into clusters of co-
cited references based on how tightly they were coupled. Each cluster is automati-
cally labeled with words from the titles of articles that are responsible for the for-
mation of the cluster. For example, clusters such as #86 scientific specialty, #76
co-citation indicator, and #67 author co-citation structure are found in the region
with many areas in blue color. The few clusters in the middle of the map connect
the upper and lower parts, including #21 cocitation map and #26 information sci-
ence. Clusters in the lower left areas are relatively new, including #37 interdisci-
plinarity and #56 visualization. Technical advances in the past 10 years have made
such visual analytics more accessible than before.

35
Fig. 5.25. A visualization of the literature of co-citation analysis.
Researchers began to realize that to capture the dynamics of science in action,
science mapping needs to bring in different perspectives and metaphors. Loet
Leydesdorff of University of Amsterdam argued that evolutionary perspectives
are more appropriate for mapping science than a historical perspective commonly
taken by citation analysts (Loet Leydesdorff & Wouters, 2000). Leydesdorff sug-
gested that the metaphor of geometrical mappings of multidimensional spaces is
gradually being superseded by evolutionary metaphors. Animations, movies, and
simulations are replacing snapshots. Science is no longer perceived as a solid
body of unified knowledge in a single cognitive dimension. Instead, science may
be better represented as a network in a multi-dimensional space that develops not
only within the boundaries of this space, but also by co-evolutionary processes
creating dimensions to this space. Now it is time to zoom closer to the map and
find trails that can lead us to the discovery of what happened in some of the most
server and long-lasting puzzle-solving cases in modern science. In the next chap-
ter, we will focus on the role of Kuhn’s paradigm shift theory in mapping scien-
tific frontiers.

36
View publication stats

You might also like