1. Analysis of Topological Characteristics of Huge Online Social Networking Services 2007.3.9. Friday 10am Telefonica Barcelona Yong-Yeol Ahn Seungyeop Han Haewoon Kwak Sue Moon Hawoong Jeong To be presented at WWW 2007

2. Online Social Networking Services Portal for people to …  Stay in touch with friends  Share photos and personal news  Find others of common interests  Establish a forum for discussion 2

3. CyWorld Largest SNS in South Korea  Started in September 2001  10 million users in 2004  16 million users out of 48 million population Front runner of many features  Friend (il-chon) relationship  Guestbook  Testimonial  Photos - scraps  Avatar in cyber home 3

4. My CyWorld “Mini-Homepage” 4

5. Overview of the Talk Snowball sampling CyWorld, MySpace, orkut data sets Metrics representative of topological characteristics Related works Analysis of CyWorld, MySpace, orkut Summary Future work 5

6. Snowball Sampling (I) Only feasible sampling method for crawling the web Select a seed node

7. Snowball Sampling (II) Pick all nodes directly connected to the seed node (1 st layer)

8. Snowball Sampling (III) Pick all nodes directly connected to 1 st layer nodes (2 nd layer)

9. CyWorld Data Sets Complete snapshot (Nov 2005)  191 million friend relationships between 12 million users  Two additional snapshots (Apr/Sep 2005) Testimonial network  100,000 users by snowball sampling 9

10. MySpace Data Set Largest in the world  Began in Jul 2003  Has 130 million by Nov 2006 Snowball sampled  During Sep/Oct 2006  Random seed to 100,000 users  About 23% of users had friend list hidden 10

11. Orkut Data Set Google SNS  Began in Sep 2002  Became official Google service in Jan 2004  Began as invitation-only; open now  Has 33 million users Snowball sampled  During Jun to Sep 2006  100,000 users 11

12. Metrics of Interest Degree distribution  “Power-law” Small number of nodes have large number of links Clustering coefficient C(k)  # of existing links / # of all possible links between a link’s adjacent neighbors Close to 1, close to a mesh Degree correlation k nn  Degree k ~ mean degree of adjacent neighbors of nodes with degree k  Assortativity: characteristic of k nn distribution 12

13. Assortative Mixing 13 M. E. J. Newman, Phys. Rev. Lett. 89, 208701 (2002) “Social” “non- social” + - degree assortative

14. Previous Works Other networks have been examined  Milgram’s seminal degree-of-separation experiment By snail mail  Movie actors, human sexual contacts, scientific collaborators  All heavy tailed and have assortative mixing Work on online SNS  Pussokram.com Internet dating community of 30,000 users  LiveJournal 1.3 million bloggers with email addr/geography  Both show some super-heavy tail end

15. Questions We Raise What are the main characteristics of online SNSs? How representative is a sample network? How does a social network evolve? 15

16. CyWorld Complete data set  Degree distribution  Clustering coefficient  Degree correlation  Average path length Historical analysis  Growth in numbers  Degree distribution, clustering coefficient, degree correlation  Average path length over time Testimonial network 16

17. Degree Distribution 17 Figure 1-(a): degree distribution, CCDF Two scaling regions

18. Clustering Coefficient Distribution 18

19. Degree Correlation 19 Not assortative

20. Average Path Length 20 < 5 is about 90%

21. Historical Analysis 21

22. Evolution of Degree Distributions 22 Two kinds of driving force

23. Evolution of Clustering Coefficient 23 Becoming more dense

24. Evolution of Degree Correlation 24 Transition from highly dissortative to slightly assortative mixing – sign of forming ‘real-world social relationship’?

25. Evolution of Path Length 25 Start of densification?

26. CyWorld Testimonials 26 # of Friends ≤ 100 In CyWorld Testimonials

27. Degree Correlation 27 Assortative Not assortative

28. Sample Networks Rare opportunity to validate snowball sampling 28

29. Degree Distributions 29

30. Clustering Coefficients 30

31. Degree Correlations 31

32. Sample Networks Degree correlation  Overestimate exponents Not fit for clustering coefficient/degree correlation 32

33. CyWorld, MySpace, and orkut Three major online SNSs of the world Common traits? 33

34. Degree Distributions 34

35. Clustering Coefficient 35 The restriction of orkut’s testmonial network system (# of friends is less than about 1000)

36. Figure 8-(c): The degree correlation of two social networks: orkut and MySpace Degree Correlation 36 Assortative “Real” Social Network Dissortative

37. Summary and Conclusions Cyworld has two scaling regions in the degree distribution.  Other measurements (C(k), degree correlation) also support the existence of two regions.  Boundary of two regions ~ 100s Dunbar’s Law: limit on human social relationships  Mature online social community orkut shows fast-decaying degree distribution and assortative mixing pattern – similar to the small degree region of Cyworld.  Close-knit, real-world-like social network structure. MySpace shows heavy tail and dissortative mixing pattern – similar to the large degree region of Cyworld.  popularity is a more important than human interaction

38. Future Work Friendship network topology not representative of activities  Identify steady core and study its evolution Explicit vs implicit communities  Clubs, towns vs cliquish behavior Growth model  Existing preferential attachment-based models do not fit  Forest fire model? Extensions? 38

39. BACKUP SLIDES 39

40. Signature of Two Scaling Regions PetterHolme, Christofer R. Edling, and Fredrik Liljeros, Structure and time evolution of an Internet dating community, SocialNetworks, 26, 155 (2004)