=Paper=
{{Paper
|id=Vol-2533/paper11
|storemode=property
|title=Jaccard Index–Based Assessing the Similarity of Research Fields in Dimensions
|pdfUrl=https://ceur-ws.org/Vol-2533/paper11.pdf
|volume=Vol-2533
|authors=Serhiy Shtovba,Mykola Petrychko
|dblpUrl=https://dblp.org/rec/conf/dcsmart/ShtovbaP19
}}
==Jaccard Index–Based Assessing the Similarity of Research Fields in Dimensions==
https://ceur-ws.org/Vol-2533/paper11.pdf
Jaccard index-Based Assessing the Similarity of
Research Fields in Dimensions
Serhiy Shtovba[0000-0003-1302-4899], Mykola Petrychko[0000-0001-6836-7843
Vinnytsia National Technical University, Khmelnytske Shose, 95, Vinnytsia, 21021, Ukraine
shtovba@vntu.edu.ua
petrychko.myckola@gmail.com
Abstract. We calculated the similarity of 10878 pairs of the research fields for
2010-2019 and for 2000-2009 on base of Dimensions data. The similarity of the
research fields is assessed by Jaccard index. The most tied research fields for
2010-2019 are Specialist Studies in Education and Curricular and Pedagogy,
whereas this pair holds rank #6 for 2000-2009. Dynamics of Jaccard index dis-
tributions does not confirm a hypothesis about increasing the share of interdis-
ciplinary research over time. It is detected that the most collaborated research
field for the both time intervals is Mechanical Engineering. This research field
has maximum sum of Jaccard indexes with all the other research fields (so
called stickiness index). It is detected that the most tied triad of research fields
for the both time intervals are Commercial Services, Marketing, and Tourism.
Distribution of similarity of the research fields, distribution of stickiness of sin-
gle research field, and distribution of stickiness of triad of the research fields
look like to Zipf law with 3 stable zones.
Keywords: scientometrics, research field, similarity, interdisciplinary, Dimen-
sions, Jaccard index, stickiness index, distribution, categorization, Zipf law.
1 Introduction
The question of how strong the various research fields are interconnected has a long
history. The relevance of the quantitative assessment of a similarity of the research
fields has intensified recently, when we are increasingly hearing about interdiscipli-
nary research, about PhD theses on boarding line of specialties, about multi-skill re-
search team etc. Traditionally, a similarity of the research fields is assessed by ex-
perts. The assessment is based on subject matters under investigation and can be asso-
ciated with an essentialist view. In [1] information-theoretic measure of linguistic
similarity to investigate the organization and evolution of scientific fields is proposed.
An analysis of almost 20M papers from the past three decades reveals that the linguis-
tic similarity is related but different from experts and citation-based classifications,
leading to an improved view on the organization of science.
Recently, citations-based approach to similarity assessment of the research fields is
most widely used. This approach is based on Porter’s metrics of paper interdiscipli-
nary [2]. The metrics is used as source data a number of citations from current paper
Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0)
2019 DCSMart Workshop.
with certain research field to papers with other research fields and vice versa. A rela-
tion of the paper to some research fields is inherited by their journal status. 2D-
distribution by research fields in axes “References to outside the research field – Cita-
tion from outside the research field” is shown in [3]. This 2D-distribution allows iden-
tifying most interdisciplinary (high stickiness) research fields.
Due to the different content of various papers of the same journal the journal-based
classification to research field produces some errors. Moreover the journal-based
classification would not allow categorize grants, patents, books etc. It is the reason to
create in [4] a new approach to paper categorization approach using machine learning
techniques. The title and abstract of the paper are source data for this categorization.
That machine learning approach for paper categorization is implemented in Dimen-
sions – the newest platform of indexing the research papers. Now, Dimensions is
worldwide biggest one with good searching services, hence it is reasonable to try
evaluating the similarities of the researches fields using Dimensions data and compare
new detected dynamics of interdisciplinary with results of previous study in [5].
2 System of research fields in Dimensions
Dimensions uses a short version of the Australian and New Zealand Standard Re-
search Classification. There are 154 fields of research and 22 domains. For example,
domain 01 – Mathematical Sciences contains the following research fields: 0101 –
Pure Mathematics; 0102 – Applied Mathematics; 0103 – Numerical and Computa-
tional Mathematics; 0104 – Statistics; 0105 – Mathematical Physics.
At the moment, Dimensions indexes 106M research papers. Each paper is assigned
to one or several research fields. Most filled research field is Clinical Sciences with
9.3M papers. Distributions of papers by the research fields for 2010-2019 and for
2000-2010 are shown in Figure 1. Both distributions look pretty much as Zipf law in
almost whole range.
Fig. 1. Distribution of number of the papers in Dimensions by research fields in semilog format
(research fields with less than 1000 publications for any of time intervals are excluded)
3 Jaccard index for two research fields
Dimensions provides a number of papers assigned with some research field. Addi-
tionally Dimensions provides number of above mentioned papers, which assigned
with each other of the research fields. It allows obtain data for calculation a similarity
of any pair of research fields using just 2 queries. For processing all the research
fields it is necessary one query per one research field.
We propose to assess a similarity of two research fields A and B by Jaccard index:
N A B
S ( A, B ) ,
N A N B N A B
where N A denotes a number of papers in research field A, N B denotes a number of
papers in research field B, N AB denotes a number of papers assigned to research
field A and to research field B simultaneously.
Consider the following example with the real data. There are 74519 papers as-
signed to Marketing in 2010 – 2019, 11843 of them also belong to Tourism. Research
field Tourism has 36274 papers, hence the similarity of Marketing and Tourism
equals to 11843 / (74519+36274-11843)=0.12.
Distributions of Jaccard index for two research fields are shown in Figure 2. For
reliability reason tiny research fields with less than 1000 papers for any of the decades
are out of consideration. In total 147 research fields are considered, hence the simi-
larity of 10878 pairs of the research fields for 2010-2019 and for 2000-2010 are cal-
culated. Most of the research fields have zero similarity. Both distributions look like
as three-zone Zipf law: the first zone for high similar research fields; the second zone
for low similar research fields; the third zone for tiny similar research fields. A strong
decreasing the similarity over ranks takes place for the first and the third zones.
Fig. 2. Distribution of the similarity of pair of the research fields (semilog format)
Figure 2 shows that the leaders for 2000-2010 have higher Jaccard indexes than
leaders for 2010-2019. But, distribution for 2010-2019 has more heavy tail. It allows
to inference by eye that there is no significant difference in change of interdisciplinary
research share over time. This conclusion concordances with a result in [5] about
science is becoming more interdisciplinary during 1975 – 2005, but in small steps.
Moreover, sum of Jaccard indexes for all the pairs of the research fields shows a
slight decreasing of interdisciplinary research share over time. This sum equals to
19.6 for 2000-2010, and equals to 18.7 for 2010-2019.
The following pairs of the research fields from top-halves of the distributions in
Figure 2 changed their Jaccard index drastically during the last decade: Environmen-
tal Engineering and Manufacturing Engineering – +254.5%; Atmospheric Sciences
and Aerospace Engineering – -67.1%.
Twenty most tied pairs of the research fields are ranked in Figure 3 and Figure 4.
Similarity of all the research fields in Figure 3 and Figure 4 are reasonable. There is
one significant change over years in the tops-20 – Specialist Studies in Education with
Curricular and Pedagogy moved from #6 for 2000-2009 to #1 for 2010-2019.
Fig. 3. Most tied pairs of the research fields for 2010-2019
Fig. 4. Most tied pairs of the research fields for 2000-2009
As an example, Figure 5 shows similarity of research fields that belong to Infor-
mation and Computer Sciences domain. There are 8 research fields (A), (B), …, (H)
in the domain. They are contoured by a blue rectangle in the Figure 5. Seven research
fields that most tied with research fields (A), (B), …, (H) are also shown on Figure 5.
Among those seven research fields Communications Technologies is most tied with
Information and Computer Sciences domain.
Let’s introduce stickiness index G(A) of a research field A as a sum of all Jaccard
indexes between A and other research fields as follows:
G( A) S ( A, p) ,
p: pF, p A
where F denotes a set of the research fields. Stickiness indexes allow to detect re-
search fields with the highest attractive to collaboration, with the highest ability to
interdisciplinary. The distributions of single research field stickiness indexes are
shown in Figure 6. Both distributions have very heavy tails, hence difference in 5-10
positions is not important.
Fig. 5. Similarity of research fields that belong to Information and Computer Sciences domain
(data for 2010-2019)
Fig. 6. Distribution of the research field stickiness indexes (semilog format)
The highest stickiness index has Mechanical Engineering for both time intervals
(Figure 7 and Figure 8). Comparing the top-rank lists, it is found that Mechanical
Engineering has lost their gap. Some significant differences in Figure 6 and Figure 8
are as follows: Environmental Science and Management jumped from #17 to #2,
Condensed Matter Physics descended from #2 to #20+. The following research fields
from the top halves of the distributions changed their stickiness indexes significantly
during the last decade:
Ecological Applications – +46.7%;
Aerospace Engineering – -37.7%;
Atmospheric Sciences – -28.2%;
Optical Physics – +27.6%;
Language Studies – -27.4%;
Astronomical and Space Sciences – -25.9%;
Genetics – +25.0%;
Environmental Science and Management – +24.6%;
Condensed Matter Physics – -24.6%;
Soil Sciences – +23.9%;
Curriculum and Pedagogy – +22.7%;
Civil Engineering – -20.5%.
Fig. 7. Most collaborative single research fields for 2010-2019
Fig. 8. Most collaborative single research fields for 2000-2009
4 Similarity of triad of research fields
A similarity of triad of research fields (A, B, C) is proposed assess by so called triad
stickiness index G( A, B, C ) as follows:
G( A, B, C) S ( A, B) S ( A, C) S ( B, C)
We found 110451 triads for 2000 – 2009 and 135503 triads for 2010 – 2019 with
non-zero stickiness index. Distributions of the triad stickiness indexes (Figure 9) and
distributions of Jaccard indexes of pairs of research fields (Figure 2) have the same
shape.
Dozens of the most tied triads are shown in Table 1 and Table 2. All the triads are
reasonable. Leader for the both time intervals is Commercial Services, Marketing and
Tourism. For this triad we constructed iteratively a map of science with nearest re-
search fields (Figure 10). On each iteration we choose a research field that maximizes
new net stickiness index. The iteration traces are present at right part of Figure 10. It
starts with Business and management, and finishes on Historical Studies. Figure 10
shows that research fields with the same domains are stronger tied each other than
with research fields from different domains. It may considerate as an indirect support
for proposed stickiness index.
Fig. 9. Distribution of triad stickiness indexes (semilog format)
Fig. 10. Map of science for the most tied triad of research fields
(edge thickness equals to Jaccard index, vertex area equals to a number of papers)
Table 1. Most tied triads for 2010–2019
Rank Research fields in the triad Stickiness index
1 Commercial Services 0.4669
Marketing
Tourism
2 Film, Television and Digital Media 0.3894
Journalism and Professional Writing
Communication and Media Studies
3 Urban and Regional Planning 0.3835
Transportation and Freight Services
Human Geography
4 Education Systems 0.3469
Curriculum and Pedagogy
Specialist Studies In Education
5 Curriculum and Pedagogy 0.3179
Specialist Studies In Education
Linguistics
6 Curriculum and Pedagogy 0.3150
Specialist Studies In Education
Sociology
7 Ecological Applications 0.3074
Environmental Science and Management
Ecology
8 Geochemistry 0.2992
Geology
Geophysics
9 Design Practice and Management 0.2949
Art Theory and Criticism
Visual Arts and Crafts
10 Civil Engineering 0.2942
Urban and Regional Planning
Transportation and Freight Services
11 Art Theory and Criticism 0.2897
Performing Arts and Creative Writing
Visual Arts and Crafts
12 Curriculum and Pedagogy 0.2834
Specialist Studies In Education
Psychology
Table 2. Most tied triads for 2000–2009
Rank Research fields in the triad Stickiness index
1 Commercial Services 0.5307
Marketing
Tourism
2 Urban and Regional Planning 0.4203
Transportation and Freight Services
Human Geography
3 Film, Television and Digital Media 0.3936
Journalism and Professional Writing
Communication and Media Studies
4 Civil Engineering 0.3593
Urban and Regional Planning
Transportation and Freight Services
5 Language Studies 0.3493
Linguistics
Literary Studies
6 Geochemistry 0.3435
Geology
Geophysics
7 Design Practice and Management 0.3293
Art Theory and Criticism
Visual Arts and Craft
8 Building 0.3216
Commercial Services
Tourism
9 Geochemistry 0.3105
Geology
Physical Geography and Environmental Geoscience
10 Architecture 0.3095
Urban and Regional Planning
Transportation and Freight Services
11 Art Theory and Criticism 0.3077
Visual Arts and Crafts
Curatorial and Related Studies
12 Environmental Science and Management 0.3028
Urban and Regional Planning
Transportation and Freight Services
5 Conclusion
For the first time similarity of the research fields in form of Jaccard index are assessed
by Dimensions data. Source information for Jaccard index calculation is the statistics
of categorized papers in Dimensions. It is calculated the similarity of 10878 pairs of
the research fields for 2010-2019 and for 2000-2009. The most tied pair of the re-
search fields for 2010-2019 is Specialist Studies in Education and Curricular and
Pedagogy, whereas they hold rank #6 for 2000-2009. The following pairs of the re-
search fields changed their Jaccard index drastically during the last decade: Environ-
mental Engineering and Manufacturing Engineering jumped on 254.5%, Atmospheric
Sciences and Aerospace Engineering dropped on 67.1%.
Stickiness index of a research field is introduced as an assessment of its attractive-
ness to be a member in an interdisciplinarity group. Stickiness index is a sum of all
Jaccard indexes related to current research field. The highest stickiness index has
Mechanical Engineering for both time intervals. The following research fields
changed their stickiness indexes drastically during the last decade: Ecological Appli-
cations jumped on 47%, Aerospace Engineering dropped on 38%.
The most tied triads of research fields are detected on base of the proposed sticki-
ness index. All the triads are reasonable. Leader among the triads is Commercial Ser-
vices, Marketing and Tourism for the both time intervals.
Possible applications of proposals similarity index is categorization of researchers,
universities, journals, which act in various research topics. Similarity index allows
taking into account tails in the research profile for more correct categorization.
References
1. Dias, L., Gerlach, M., Scharloth, J., Altmann, E.G.: Using text analysis to quantify the
similarity and evolution of scientific disciplines. Royal Society Open Science. 5, (2018).
doi: 10.1098/rsos.171545.
2. Porter, A.L., Cohen, A.S., Roessner, J.D., Perreault, M.: Measuring researcher interdisci-
plinarity. Scientometrics. 72, 117–147 (2007). doi: 10.1007/s11192-007-1700-5.
3. Noorden, V.R.: Interdisciplinary research by the numbers. Nature. 525, 306-307 (2015).
doi: 10.1038/525306a.
4. Herzog, C., Kierkegaard, B.L.: Response to the letter “Field classification of publications
in Dimensions: a first case study testing its reliability and validity”. Scientometrics. 117.
641-645 (2018). doi: 10.1007/s11192-018-2854-z.
5. Porter, A., Rafols, I.: Is science becoming more interdisciplinary? Measuring and mapping
six research fields over time. Scientometrics. 81, 719-745 (2009). doi: 10.1007/s11192-
008-2197-2.