Visualization and Data Analysis 2009 San Jose, California, USA 19 January 2009

1. Understanding outside collaborations of the Chinese Academy of Science using Jensen-Shannon divergence • Visualization and Data Analysis 2009 • San Jose, California, USA • 19 January 2009 Russell J. Duhon rduhon@indiana.edu Cyberinfrastructure for Network Science Center School of Library and Information Science Indiana University Bloomington, IN, USA Russell J. Duhon: Understanding Outside Collaborations 01/28

2. Table of Contents • Collaboration • Entropy • Collaboration and Entropy • Information • Insight (Perhaps) • Applications in the Small and Large Russell J. Duhon: CIShell, Network Workbench, and the Scholarly Database 02/28

3. Collaboration What is Collaboration? Russell J. Duhon: Collaboration 03/28

4. Collaboration What is Collaboration? Good question . . . but a lot of collaboration is pretty easy to recognize. Co-authorship is a common sort looked at in scientometrics. Russell J. Duhon: Collaboration 04/28

5. Collaboration How do we measure collaboration? In a lot of ways, some of which we’ve heard about earlier today. Returning to co-authorship, one way is to make networks and analyze those. Russell J. Duhon: Collaboration 05/28

6. Collaboration What if we don’t have a nice co-authorship network? In other words, how do we uncover less obvious characteristics of interest? Russell J. Duhon: Collaboration 06/28

7. Entropy Russell J. Duhon: Entropy 07/28

8. Entropy What if we don’t have a nice co-authorship network? In other words, how do we uncover less obvious characteristics of interest? Russell J. Duhon: Entropy 08/28

9. Entropy Many conceptualizations, and in one sense helps measure distinguishability. Russell J. Duhon: Entropy 09/28

10. Entropy Shannon Entropy Russell J. Duhon: Entropy 10/28

11. Entropy Jensen-Shannon Divergence where H is the Shannon Entropy. Russell J. Duhon: Entropy 11/28

12. Collaboration and Entropy So, given how a set of collaborators co-author with some control set, we can treat those as distributions and uncover the distinguishability of those distributions. Russell J. Duhon: Collaboration and Entropy 12/28

13. Collaboration and Entropy Then, given a pairwise distinguishability metric, algorithms taking advantage of metric similarity become available. Russell J. Duhon: Collaboration and Entropy 13/28

14. Collaboration and Entropy Take those pair-wise metrics, and do Agglomerative Hierarchical Clustering with Ward’s Algorithm to look for structure. Russell J. Duhon: Collaboration and Entropy 14/28

15. Information Russell J. Duhon: Information 15/28

16. Information Ignoring Beijing-only collaborators, there are two primary clusters: Russell J. Duhon: Information 16/28

17. Information More developed, more politically normalized countries, Russell J. Duhon: Information 17/28

18. Information and less developed, less politically normalized countries. Russell J. Duhon: Information 18/28

19. Information Within the cluster of more developed, more politically normalized countries, there are again two clear clusters: Russell J. Duhon: Information 19/28

20. Information Larger, more developed countries, including the up-and-coming BRIC members, Russell J. Duhon: Information 20/28

21. Information and much smaller countries, for the most part Russell J. Duhon: Information 21/28

22. Information Looking more closely, there are many small clusters that show expected patterns. Russell J. Duhon: Information 22/28

23. Insight (Perhaps) Starting to approach insight, the locations of Chinese collaborators hint at potential explanations. Russell J. Duhon: Insight (Perhaps) 23/28

24. Insight (Perhaps) ? , , One of the strongest potential applications will be identifying possible sites for the application of policy tools to strengthen, sustain, and transform collaborative structures. Russell J. Duhon: Insight (Perhaps) 24/28

25. Applications in the Small and Large Remember the motivation: understanding structure when there is less information than ideal. Russell J. Duhon: Applications in the Small and Large 25/28

26. Applications in the Small and Large Small research groups generally know who their members collaborate with and how much, but rarely how those people collaborate with each other. Russell J. Duhon: Applications in the Small and Large 26/28

27. Applications in the Small and Large Large organizations often have similar information limitations, particularly due to identification and data dirtiness. Russell J. Duhon: Applications in the Small and Large 27/28

28. Applications in the Small and Large In both cases, this procedure may be very useful. The insights hint at potential relations among outside collaborators, and comparison of distributions minimizes the impact of dirty data. Russell J. Duhon: Applications in the Small and Large 28/28

29. Thank You Russell Duhon rduhon@indiana.edu NSF IIS-0534909 and IIS-0513650 awards