Automatic Classification and Analysis of Interdisciplinary Fields in Computer Sciences

Transcription

1 Automatic Classification and Analysis of Interdisciplinary Fields in Computer Sciences Tanmoy Chakraborty Google India PhD Fellow Indian Institute of Technology, Kharagpur India In collaboration with: Srijan Kumar, M Dastagiri Reddy, Suhansanu Kumar, Niloy Ganguly, Animesh Mukherjee IIT-Kgp, India 2013 ASE/IEEE SocialCom, Washington D.C., USA, September 8-14, 2013

2 Outline o Problem Definition o Dataset o Indicators of Interdisciplinarity o Unsupervised Classification Model o Evolution Landscape of Interdisciplinarity o Core-periphery Analysis o Conclusion

3 How to seek a good toolkit? Stitching machine Saw Paperpunch Screw-driver Bottle Opener Knife Nail-cutter COMBO PAC

4 Interdisciplinarity Chemistry Physics Engineering Economics Sociology Biology Mathematics Interdisciplinary toolkit

5 In The Lines of Great Thinkers We are not students of some subject matter, but students of problems. And problems may cut right across the borders of any subject matter or discipline. Karl Popper Interdisciplinary research is the only way to do research in current times. Fritjof Capra

6 Outline Problem Definition Dataset Indicators of Interdisciplinarity Unsupervised Classification Model Evolution dynamics of Interdisciplinarity Core-periphery Analysis Conclusion

7 Problem Definition o Proper quantitative indicators of Interdisciplinarity o Unsupervised classification of core and interdisciplinary fields o Evolution dynamics of interdisciplinarity o Core-periphery analysis of citation network

8 Outline Problem Definition Dataset Indicators of Interdisciplinarity Unsupervised classification model Evolution dynamics of Interdisciplinarity Core-periphery Analysis Conclusion

9 Dataset o Large DBLP dump used by Chakraborty et al. (ASONAM, 13) o Bibliographic information during Publicly available: Paper name - Author(s) - Publication venue - Year of publication - Abstract - References - Field 24 Fields # of valid papers 702,973 # authors 495,311 # unique venue name 1,705 AI Bioinformatics NLP Algorithm Graphics WWW Networking Comp. Vision Education Database Data Mining OS Dist Comp. Prog. Lang. Embedded Sys. Architecture Security Simulation Software Engg. IR HCI Machine Learning Scientific Comp. Multimedia

10 Citation network o Aggregated Network: Node (paper) Link (citation) o Time-stamp wise Networks: 5 years sliding window (60-64, 61-65, 62-66,..., )

11 Outline Problem definition Dataset Indicators of Interdisciplinarity Unsupervised classification model Evolution dynamics of interdisciplinarity Core-periphery organization Conclusion

12 Indicators of Interdisciplinarity o Reference Diversity Index (RDI) o Citation Diversity Index (CDI) o Membership Diversity Index (MDI) o Attraction Index Most of the indices are Entropy based measures More Entropy => More diversity

13 Reference Diversity index (RDI) RDI of a paper X i = RDI( X i ) = j p j log p j p j = proportion of references of X i citing the papers of field F j More RDI, more interdisciplinarity F j p j = 3/5 X i RDI(X i ) = - 3/5 log (3/5) 2/5 log (2/5) F k p k = 2/5 = 0.67

14 Citation Diversity Index (CDI) o CDI of a paper X i at time t i = CDI t i ( X i ) = j p j log p p j = proportion of citations received by X i from the papers of field F j p j F X i j F j Drift: CDI(X i ) = 0.67 p k F k Drift of CDI between two successive time windows = t ( f i ) = CDI t ( f ) ( 1 i CDI t f i i i + i )

15 Spikes in CDI

16 Membership Diversity Index (MDI) 1. Identify overlapping communities [Xie et al., ICDM, 2011] AI IR 2. Tag the communities by the fields (major field in a group) 3. For each field f i in the dataset, 3.1 Observe the belongingness of all papers in different field-tagged communities 3.2 Measure MDI MDI ( f i ) = j p j log where, p j is the fraction of overlapped papers of field f i belonging to the communities tagged as f j More MDI, more interdisciplinarity p j

17 External Evidence: Attraction index External Evidence: χ f n = i + 4 c i n i n i : # unique authors up to the year t i (in field f ) n i+4 : # unique authors up to the year t i+4 (in field f ) c i : # publications in f in the time window (t i+4 - t i ) More χ, more interdisciplinarity 0 n i = 2 t i t i+4 n i+4 = 3 χ f =1/4

18 Outline Problem definition Dataset Indicators of Interdisciplinarity Unsupervised Classification Model Evolution dynamics of interdisciplinarity Core-periphery organization Conclusion

19 Unsupervised Classification Model o A field is represented by a vector of size 4 indicating four features o Adjacency matrix A of size A(i,j)= Cosine similarity of field i and j o Clustering algorithm proposed by Waltman et al. (J. Informetrics, 2010)

20 Result of the Classification

21 Outline Problem definition Dataset Indicators of Interdisciplinarity Unsupervised classification model Evolution Landscape of Core-periphery organization Conclusion Interdisciplinarity

22 Evolution dynamics o Construct a field-field citation network o 24 nodes in each time-stamp o Draw directed and weighted edges based on citations o Observe the citation distribution across the fields

23 Evolutionary Landscape PL WWW o Fields are grouped based on the connection proximity o The size of the font indicates the relative importance (# of incoming citations) of a field

24 Outline Problem definition Dataset Indicators of Interdisciplinarity Unsupervised classification model Evolution dynamics of interdisciplinarity Conclusion Core-periphery Analysis

25 Core-periphery Analysis ALGO ALGO ALGO ALGO DM DM DM DM

26 Outline Problem definition Dataset Indicators of Interdisciplinarity Unsupervised classification model Evolution dynamics of interdisciplinarity Core-periphery organization Conclusion

27 Conclusions o Quantitative indications of interdisciplinary o Automated scheme to identify interdisciplinary fields o Evolutionary landscape depicts the cross-hybridization among fields o K-core analysis shows the steady movements of interdisciplinary field at the core Future Directions: o Identify interdisciplinary papers o Predicting the possible fields to be intermingled next.

28 Acknowledgements o Financial Support: Google India Pvt. Ltd. o Technical support: Complex Network Research Group (CNeRG), IIT-Kgp

29 Conclusion o Quantitative indications of Interdisciplinarity o Citation based indices (i) suggests the practice of interdisciplinairty (ii) unfolds the evolutionary landscape of Interdsiciplinarity o Few interdisciplinary fields (Data Mining) come towards the core o Future work: Identification of interdisciplinary papers Recommendation system to predict future combination of fields