“The Rise and Rise of Citation Analysis ”

Post on 22-Feb-2016

94 views 0 download

Tags:

description

“The Rise and Rise of Citation Analysis ”. Tanmoy Chakraborty CNeRG , IIT- Kgp , India. Twofold Research Interests. Analyzing communities/clusters in complex networks Studying different aspects of citation network . “The Rise and Rise of Citation Analysis ”. In collaboration with - PowerPoint PPT Presentation

Transcript of “The Rise and Rise of Citation Analysis ”

“The Rise and Rise of Citation

Analysis”Tanmoy ChakrabortyCNeRG, IIT-Kgp, India

Twofold Research Interests

• Analyzing communities/clusters in complex networks

• Studying different aspects of citation network

“The Rise and Rise of Citation

Analysis”In collaboration with

Suhansanu Kumar, Pawan Gowel, Animesh Mukherjee, Niloy Ganguly

Mixed Sentiment • Sense and non-sense about citation analysis (*6860)

--- T. Opthof, Cardiovascular research, 97

• The rise and rise of citation analysis (*1399) --- Lokman I. Meho, Phy. Res., 07

• Does citation pay? (*887) --- Fowler & Aksnes, Scientometrics, 07

• Think beyond citation analysis (*1009) --- Sarli et al., NIPS, 10

Raw Citations Count To assess

• Quality of a paper• Prominence of a researcher• Success of a collaboration/group• Quality of a conference/journal• Quality of an Institute• Impact of a research area

Only Citation Count

Sooner or later, you will definitely be subjected to such an analysis

Bibliometrics: Raw Citation Count

• Journal Impact factor • Immediacy factor• Eigen factor• Altmetric• 5 years Impact factor

Cita

tion

Common assumption

Publication Universe• Crawled entire Microsoft Academic Search • Papers only in Computer Science domain• Basic preprocessing

Basic Statistics of papers from 1960-2010

Values

Number of valid entries 3,473,171

Number of authors 1,186,412

Number of unique venues 6,143

Avg. number of papers per author 5.18

Avg. number of authors per paper 2.49

Publication UniverseAvailable Metadata for each paper

Title

Unique ID

Named entity disambiguated authors’ name

Year of publication

Named entity disambiguated publication venue

Related research field(s)

References

Keywords

Abstract

Citation context

Citation ProfileAn exhaustive analysis of the citation profiles

• Papers having at least 10 yrs history and consider at most 20yrs history

• Scale the entries of the citation profile between 0-1• Use peak-detection heuristics

» Each peak should be at least 75% of the max peak» Two consecutive peak should be separated at least 3 yrs

Five Universal Citation Profiles

Q1 and Q3 represent the first and third quartiles of the data points respectively.

Avg. behavior

Another category: ‘Oth’ => having less than one citation (on avg) per year

Five Universal Citation ProfilesA deeper look

Immediate Questions• Is the Journal Impact factor (JIF) formula

correct? • JIF at year 2000 : Eugene Garfield (1975)

# of citations received in 2000 by papers published in that journal in 1998 and 1999

divided by

# of papers published in the journal in 1998 and 1999

Immediate Questions

What does JIF really imply?

Importance of the recent papers in current time period?

Relevance of the journal itself in current time period

Why last 2 years?Why not all the citations at current time

Immediate Questions

Over-consider

Under-consider

JIF overlooks the importance of Peak_Late and MonIncrOver-consider

More on the Categories1. Are they biased on the year of publication?

(Aging factor)

Mean year Year deviation

Peak_Int 1994 5.19

Peak_Mul 1992 6.68

Peak_Late 1992 6.54

MonDec 1994 5.43

MonIncr 1993 5.36

Same ageAns: No

More on the Categories2. Are they biased on Journals/conferences?

Peak_Int Peak_Mul Peak_Late MonDec MonIncr% of

conferences paper 65 39.03 39.89 60.73 25.26% of

Journal paper 35 60.97 60.11 39.27 74.74

More on the Categories3. Are they affected by self-citation?

Peak_Int Peak_Mul Peak_Late MonDec MonIncr Oth

Peak_Int 0.72 0.10 0.03 0.01 0 0.15

Peak_Mul 0.02 0.81 0.04 0 0.01 0.11

Peak_Late 0.01 0.06 0.86 0 0.01 0.06

MonDec 0.05 0.14 0 0.41 0 0.35

MonIncr 0 0.02 0.01 0 0.88 0.09

Transition matrix showing the transition of categories after removing self citations

Most affected by self citation

Least affected by self citations

More on the Categories4. What about Peak_Mul?

Peak_Mul Might be Intermediary between Peak_Int and Peak_Late

2.5

5.3

5.1

3.1

4.2

5.3 12.1

10.8

Time

Avg

Pea

k H

eigh

t

Peak_Int

Peak_Mul

Peak_Late

Years after publication

3.1+2.5 = 5.6

(12.1-5.3) = 6.8 ~ (10.8-4.2)

Where does this classification help?

• To improve Bibliometrics in scientific research• Various prediction systems

• Future citation prediction system• Predicting emerging field/topic• Predating future star/seminal papers

• Paper search and Recommender systems

On predicting

Future Citation Count at the Time of Publication

Problem Definition

Traditional FrameworkYan et al., JCDL, 2012

Problems in Traditional Framework• Consider initial few years’ statistics after publication

Proved to be very effective

• Lack of time dimension in prediction

• Suffers a lot from outlier points during regression

Problems in Traditional Framework:how to tackle

• Consider initial few years’ statistics after publication Proved to be very effective

• Lack of time dimension in prediction

• Suffers a lot from outlier points during regression

Try to predict citations as early as possible(may be at the time of publication )

Should consider the time dimension

Reduce outlier points as much as possible

Our Framework: 2-stage Model

FeaturesAuthor-centric

Productivity (Max/Avg)

H-index (Max/Avg)

Versatility (Max/Avg)

Sociality (Max/Avg)

Performance Evaluation(i) Coefficient of determination (R2)

The more, the better

(ii) Mean squared error (θ) The less, the better

(iii) Pearson correlation coefficient (ρ) The more, the better

Performance of SVMConfusion Matrix

Performance of Regression Model

Feature Analysis

• Five universal citation profiles• Different analysis on these categories• Can help to reframe existing bibliometrics• Can be a generic way in machine learning• Can enhance the performance of the existing

systems

Conclusion

Thank You