Ranking influential communities in networks · David Selby is supported by EPSRC grant EP/M508184/1...
1
biomedical sciences chemistry & physics medicine zoology economics mathematics informatics statistics rheumatology textiles 0.5 1.0 1.5 2.0 1st 20th 40th 60th 80th 100th biomedical sciences chemistry & physics medicine neuroscience zoology geology economics mathematics psychology oncology informatics psychiatry parasitology environmental science agriculture surgery politics management sociology material science education astrophysics food science statistics energy radiology spectroscopy orthopaedics linguistics human geography particle physics veterinary medicine nephrology ecology biomaterials obstetrics operational research toxicology dentistry sports medicine opthalmology rheumatology complexity robotics rhinology marketing electronics law reviews social history anthropology urology communication dermatology logistics electrical engineering information systems bibliometrics tribology history of science computer graphics social anthropology psychometrics plastic surgery learning disabilities policy geotechnology social work international law logic philosophy vascular surgery social justice radioactivity tourism forensics mycology transfusion engineering management mathematical finance measurement Pharm Stats Stats in Medicine Pakistan J Stats Comp Stats ANZ J Stats Comm Stats: Sim & Comp Comp Stats & Data Analysis J Applied Stats JRSS-A JRSS-C J Time Series Appl Stoch Models in Business J Machine Learning Res J Multivariate Analysis Machine Learning JASA Stats & Comp Bernoulli Methods & Comp in Appl Prob J Agr Bio & Envir Stats Brazil J Prob & Stats Comm Stats: Theory & Methods Stats Papers Stats & Prob Letters Statistics Test J Korean SS Rev Colomb Estad J Stat Plan & Infer Ann Inst Stat Maths J Stats Comp & Sim Annals Biometrika EJ Stats ESAIM Prob & Stats Extremes JRSS-B Metrika Prob Eng & Inf Sci Scand Actuarial J Scan J Stats Seq Analysis Stoch Models Adv Data Analysis & Class J Stat Software Stat Methods & Appl Qual Rel Eng Int Technometrics Biometrics Stats in Biopharm Res Ann Appl Stats Stat Sci Environometrics Stat Meth Med Res Bayesian Analysis Biometrical J Biostatistics Int J Biostatistics J Biopharm Stats Lifetime Data Analysis Stats & Interface American Stat Envir & Eco Stats Int Stats Rev JCGS Stat Method Stat Model Canada J Stats J Nonpar Stats J Qual Tech Adv Stat Analysis REVSTAT Qual Tech & Quant Mgmt R Journal Stat Neerlandica SORT Qual Eng Statistics 77 journals 29,000 citations Ranking influential communities in networks Ranking journals How to measure influence using random walks The PageRank algorithm models the behaviour of a typical PhD student. 1. Open a random journal. 2. Pick a random reference and open the cited journal. 3. Repeat, ad nauseum. PageRank is the proportion of time spent reading each journal, i.e. the stationary distribution of an ergodic Markov chain. It is a measure of total influence. PageRank has a size bias: bigger journals have more/longer articles in them, attracting more citations. What if we want to measure prestige, rather than popularity? The Scroogefactor score, defined as PageRank per reference, controls for this size bias. It measures influence weight per outgoing citation. Like the Bradley–Terry model, journals are, in effect, penalised for being generous with citations and rewarded for being miserly. When the Bradley–Terry model fits exactly, a journal's Scroogefactor is exactly equal to its Bradley–Terry score. What can citation data reveal about the flow of influence in scientific fields? Web of Science 10,713 journals 112 communities 9.1 million citations Finding communities How to group journals into fields using random walks The Infomap algorithm (Rosvall & Bergstrom 2008; PNAS 105) looks for a clustering of nodes that gives the shortest possible description of a random walk around the network. Every node has a twolevel address, identifying a community and the node's position within that community. It is like a street address, comprising a street name (position) and city name (group). Street names are often reused between cities, but every street–city pair is unique. By aggregating the journals in each community into a single ‘superjournal’, we can model the exchange of citations between disciplines. Acknowledgements David Selby is supported by EPSRC grant EP/M508184/1 and David Firth is supported by the EPSRC ilike programme. Thanks to Thomson Reuters for providing Journal Citation Reports data in a convenient format. R code and full output from the analysis is available on GitHub: https://github.com/Selbosh/user2017 Above: a betweenfields ranking. Applied, general disciplines tend to have higher Bradley–Terry ability scores than highly specialised or theoretical ones. Below: a withinfield ranking for statistics journals. A high score implies the journal is influential within statistics, but not necessarily influential on other fields. JRSS-A American Stat JRSS-B Annals Biometrika JCGS Machine Learning Stats in Medicine Metrika Statistics R Journal 1 2 1st 20th 40th 60th Statistical model Given a set of paired comparisons, the Bradley–Terry model estimates an ability score for each object, such that for any pair of objects i and j. Citations between academic journals can be treated as paired comparisons: being cited means being an ‘exporter of intellectual influence’ (Stigler 1994; Statistical Science). Using ability scores, we can predict the probability that journal i cites journal j more than j cites i. Influential journals are more likely to be cited by other influential journals. P(i beats j ) 1 - P(i beats j ) = μ i μ j Right: the Web of Science network. Each node represents a community of journals. The edges represent interdisciplinary citations. Below: journals and citations within the statistics community. David Selby [email protected] David Firth [email protected]
Transcript of Ranking influential communities in networks · David Selby is supported by EPSRC grant EP/M508184/1...
biomedical sciences
chemistry & physics
medicine
zoology
economics
mathematicsinformatics
statisticsrheumatology
textiles0.5
1.0
1.5
2.0
1st20th40th60th80th100th
biomedical sciences
chemistry & physics
medicine
neuroscience
zoologygeology
economics
mathematics
psychology
oncology
informatics
psychiatry
parasitology
environmental science
agriculture
surgery
politics
management
sociology
material science
education
astrophysics
food science
statistics
energy
radiology
spectroscopy
orthopaedics
linguistics
human geography
particle physics
veterinary medicine
nephrology
ecology
biomaterials
obstetrics
operational research
toxicologydentistry
sports medicine
opthalmology
rheumatology
complexity robotics
rhinology
marketing
electronics
law reviews
social history
anthropology
urology
communication
dermatology
logistics
electrical engineering
information systems
bibliometrics
tribologyhistory of science
computer graphics
social anthropology
psychometrics
plastic surgery
learning disabilities
policy
geotechnology
social work
international law
logic
philosophy
vascular surgery
social justice
radioactivity
tourism
forensics
mycologytransfusion
engineering management
mathematical finance
measurement
Pharm Stats
Stats in Medicine
Pakistan J Stats
Comp Stats
ANZ J Stats
Comm Stats: Sim & Comp
Comp Stats & Data Analysis
J Applied Stats
JRSS-A
JRSS-C
J Time Series
Appl Stoch Models in Business
J Machine Learning Res
J Multivariate Analysis
Machine Learning JASA
Stats & Comp
Bernoulli
Methods & Comp in Appl Prob
J Agr Bio & Envir StatsBrazil J Prob & Stats
Comm Stats: Theory & Methods
Stats Papers
Stats & Prob Letters
Statistics
Test
J Korean SS
Rev Colomb Estad
J Stat Plan & Infer
Ann Inst Stat Maths
J Stats Comp & Sim
Annals
Biometrika
EJ Stats
ESAIM Prob & Stats
ExtremesJRSS-B
Metrika
Prob Eng & Inf SciScand Actuarial J
Scan J Stats
Seq Analysis
Stoch Models
Adv Data Analysis & Class
J Stat Software
Stat Methods & Appl
Qual Rel Eng Int
Technometrics
Biometrics
Stats in Biopharm Res
Ann Appl Stats
Stat Sci
Environometrics
Stat Meth Med Res
Bayesian Analysis
Biometrical J
Biostatistics
Int J Biostatistics
J Biopharm Stats
Lifetime Data Analysis
Stats & Interface
American Stat
Envir & Eco Stats
Int Stats Rev
JCGS
Stat Method
Stat Model
Canada J Stats
J Nonpar Stats
J Qual Tech
Adv Stat Analysis
REVSTAT
Qual Tech & Quant Mgmt
R Journal
Stat Neerlandica
SORT
Qual Eng
Statistics77 journals29,000 citations
Ranking influential communities in networks
Ranking journalsHow to measure influence using random walks
The PageRank algorithm models the behaviour of a typical PhD student.
1. Open a random journal.
2. Pick a random reference and open the cited journal.
3. Repeat, ad nauseum.
PageRank is the proportion of time spent reading each journal, i.e. the stationary distribution of an ergodic Markov chain. It is a measure of total influence.
PageRank has a size bias: bigger journals have more/longer articles in them, attracting more citations. What if we want to measure prestige, rather than popularity?
The Scroogefactor score, defined as PageRank per reference, controls for this size bias. It measures influence weight per outgoing citation.
Like the Bradley–Terry model, journals are, in effect, penalised for being generous with citations and rewarded for being miserly. When the Bradley–Terry model fits exactly, a journal's Scroogefactor is exactly equal to its Bradley–Terry score.
What can citation data reveal about the flow of influence in scientific fields?
Web of Science10,713 journals
112 communities9.1 million citations
Finding communitiesHow to group journals into fields using random walks
The Infomap algorithm (Rosvall & Bergstrom 2008; PNAS 105) looks for a clustering of nodes that gives the shortest possible description of a random walk around the network.
Every node has a twolevel address, identifying a community and the node's position within that community. It is like a street address, comprising a street name (position) and city name (group). Street names are often reused between cities, but every street–city pair is unique.
By aggregating the journals in each community into a single ‘superjournal’, we can model the exchange of citations between disciplines.
AcknowledgementsDavid Selby is supported by EPSRC grant EP/M508184/1 and David Firth is supported by the EPSRC ilike programme.
Thanks to Thomson Reuters for providing Journal Citation Reports data in a convenient format.
R code and full output from the analysis is available on GitHub:https://github.com/Selbosh/user2017
Above: a betweenfields ranking. Applied, general disciplines tend to have higher Bradley–Terry ability scores than highly specialised or theoretical ones.
Below: a withinfield ranking for statistics journals. A high score implies the journal is influential within statistics, but not necessarily influential on other fields.
JRSS-AAmerican StatJRSS-B
AnnalsBiometrika
JCGSMachine Learning
Stats in Medicine
MetrikaStatistics
R Journal
1
2
1st20th40th60th
Statistical modelGiven a set of paired comparisons, the Bradley–Terry model estimates an ability score for each object, such that
for any pair of objects i and j.
Citations between academic journals can be treated as paired comparisons: being cited means being an ‘exporter of intellectual influence’ (Stigler 1994; Statistical Science).
Using ability scores, we can predict the probability that journal i cites journal j more than j cites i. Influential journals are more likely to be cited by other influential journals.
P(i beats j)1− P(i beats j)
=µi
µj
Right: the Web of Science network. Each node represents a community of journals. The edges represent interdisciplinary citations.
Below: journals and citations within the statistics community.
David [email protected]
David [email protected]
