Friendship Maintenance and Prediction in Multiple Social Networks

Due to the proliferation of online social networks (OSNs), users find themselves participating in multiple OSNs. These users leave their activity traces as they maintain friendships and interact with users in these OSNs. In this work, we analyze how users maintain friendship in multiple OSNs by studying users who have accounts in both Twitter and Instagram. Specifically, we study the similarity of a user's friends and the evenness of friendship distribution in multiple OSNs. Our study shows that most users in Twitter and Instagram prefer to maintain different friendships in the two OSNs, keeping only a small clique of common friends in across the OSNs as illustrated by the left skewed friendship similarity distribution. Based upon our empirical study, we conduct link prediction experiments to predict missing friendship links in multiple OSNs using the neighborhood features, neighborhood friendship maintenance features and cross-link features. Our link prediction experiments shows that unsupervised methods can yield good accuracy in predicting links in one OSN using another OSN data and the link prediction accuracy can be further improved using supervised method with friendship maintenance and others measures as features.

Transcript

1. Friendship  Maintenance  and  Predic0on  in  Mul0ple   Social  Networks   Presented  by:  Roy  LEE   Roy  Ka-­‐Wei  LEE  &  Ee-­‐Peng  LIM  
2. Users  on  Mul0ple  Social  Networks   •  2014  survey1:  52%  of  online  users  use  2  or  more   online  social  network  sites  (OSNs)   •  New  dynamics  in     friendship   maintenance  in   mulMple  OSNs   2   1Pew  Research  Center.  Social  Media  Update  2014.  2015.  
3. Maintain  Friendship  in  Mul0ple  OSNs   3   Twi3er   Maintain  different  groups  of  friend  in  different  OSNs   for  different  purposes   Instagram  
4. 4   Twi3er   Maintain  high  overlap  of  friends  in  mulMple  OSNs  for   ease  of  maintenance   Instagram   Maintain  Friendship  in  Mul0ple  OSNs  
5. 5   Twi3er   Maintain  uneven  number  of  friends  in  different   OSNs,  keeping  only  a  smaller  group  of  close  friends   overlapped  in  mulMple  OSNs   Instagram   Maintain  Friendship  in  Mul0ple  OSNs  
6. Research  Ques0ons   1.  How  do  users  maintain  friendship  across  mulMple   OSNs?     a.  Do  they  maintain  overlap  of  friends  across  different   OSNs?     b.  Do  they  maintain  even  number  of  friends  across   mulMple  OSNs?   2.  Are  we  able  to  predict  friendship  (link)  in  an  OSN   using  the  knowledge  of  friendship  in  another   OSN?   6  
7. 7   FRIENDSHIP  MAINTENANCE  MEASURES  
8. •  Measure  the  overlap  or  similarity  of  user’s   friendship  across  mulMple  OSNs   •  Adapted  the  D-­‐Correla)on  approach  (Berlingerio  et   al.,  2011)   Friendship  Similarity     8   Denotes  a  set  of  OSNs   {N1 , N2 ,…, Nn }   Common  friends  of   user  x  across  all  OSNs   All  unique  friends  of   user x across  all  OSNs  
9. •  Define  evenness of user’s friendship distribution  in   mulMple  OSNs: Friendship  Evenness     9   RaMo  of  friends  of  a   user  x  in  OSN  Ni   relaMve  to  all  friends   Expected  raMo  of   friends  adds  to     each  OSNs  
10. 10   EMPIRICAL  STUDY  ON     TWITTER  AND  INSTAGRAM  
11. Dataset   11   97,978   Base  Users   24~  Million     Instagram  Friends   17~  Million     TwiSer  Friends  
12. User  Friend  Matching   •  There  is  a  need  to  match  friends  account  in  two   OSNs.  Problem:  few  friends  declared  both  OSN   accounts   •  Three  levels  user  friend  matching  methods:   1.  Self-­‐Repor0ng  Matching.  Friends  how  declared  both   OSN  accounts.   2.  Username  matching.  Match  by  friends’  exact   username.     3.  Username  Bigram  Matching.  An  approximate  method   that  match  friends  using  username  bigram.  To  cater   for  situaMon  for  unviability  of  usual  username.   12  
13. User  Friend  Matching  Results   13  
14. •  Lef-­‐learning  bell  shape  distribuMon  –  few  users  maintain   similar  friendship  in  their  Twiher  and  Instagram  accounts   Friendship  Similarity  Distribu0on   14  
15. •  Right-­‐learning  bell  shape  distribuMon  –  most  users  prefer  to   have  even  friendship  counts  in  different  OSNs   Friendship  Evenness  Distribu0on   15  
16. •  Few  users  maintain  similar  friendship  in  their   Twiher  and  Instagram  account   •  Most  users  prefer  to  not  have  overly  uneven   friendship  counts  in  different  OSNs   •  Possible  reason:  Users  use  different  OSNs  for   different  purpose  or  interest  (Lim  et  al.,  2015)  ,   which  indirectly  moMvates  the  users  to  connect  to   different  friends  in  different  OSNs   Empirical  Study  Summary   16  
17. 17   FRIENDSHIP  LINK  PREDICTION  
18. •  PredicMon  Tasks   o Twi:er  Link  Predic)on  (TWLP)  –  Predict  if  two  users  are   friends  in  Twiher   o Instagram  Link  Predic)on  (INLP)  –  Predict  if  two  users   are  friends  in  Instagram   •  Unsupervised  Link  PredicMon   •  Supervised  Link  PredicMon   Predic0on  Overview   18  
19. •  Dataset:  5,000  posiMve  and  25,000  negaMve  instances   •  Neighborhood  Measures:   o  Common  Neighbors  (CN)   o  Jaccard  Coefficient  (JC)   o  Adamic-­‐Adar  (AA)   •  Performance  EvaluaMon:   Unsupervised  Link  Predic0on   19   5000
20. Unsupervised  Link  Predic0on  Results   20   F1  =  0.882   F1  =  0.838  
21. •  Training  and  TesMng  Datasets:  5,000  posiMve  and  25,000   negaMve  instances       Supervised  Link  Predic0on   21  
22. •  Training  and  TesMng  Datasets:  5,000  posiMve  and  25,000   negaMve  instances       •  Neighborhood  features  from   measures  used  in  unsupervised   link  predicMon  methods   Supervised  Link  Predic0on   22  
23. •  Training  and  TesMng  Datasets:  5,000  posiMve  and  25,000   negaMve  instances       Supervised  Link  Predic0on   23   •  Common  Neighbor  Friendship   Maintenance  Features     incorporates  the  friendship   maintenance  behavior  into   common  neighbor  method    
24. •  Training  and  TesMng  Datasets:  5,000  posiMve  and  25,000   negaMve  instances       Supervised  Link  Predic0on   24   •  Cross  network  features   which  returns  1  if  the  users   of  the  instances  are  friends   in  another  network,  and  0   otherwise  
25. •  6  different  feature  configuraMons  in  our  supervised   link  predicMon  methods:   1.  NBO:  Neighborhood  features  only   2.  NFM:  Common  Neighbor  Friendship  Maintenance   features  only   3.  NBOFM:  Neighborhood  and  Common  Neighbor   Friendship  Maintenance  features   4.  NBCL:  Neighborhood  and  Cross  Network  features   5.  NFMCL:  Common  Neighbor  Friendship  Maintenance   and  Cross  Network  features   6.  ALL:  All  features   Supervised  Link  Predic0on   25  
26. •  Most  feature  configuraMons   have  F1  higher  than  best  F1   scores  of  unsupervised   methods   •  All  features  outperforms   other  methods   •  The  addiMon  of  Cross   Network  (CL)  features   improve  the  results  of  NFM   and  NBO  features   Supervised  Link  Predic0on   26  
27. •  Common  Neighbor   Friendship  Maintenance   (NFM)  features  performed   slightly  worse  than   Neighborhood  (NBO)   features   •  Possible  reason:  lack  of   common  neighbors  with   friendship  maintenance   measures  who  are  also  base   users   Supervised  Link  Predic0on   27  
28. •  Re-­‐examined    the  supervised   link  predicMon  results  and   determine  the  accuracy  of   test  instances  that  have  at   least  one  common  neighbor   who  is  also  a  base  user   •  Common  Neighbor   Friendship  Maintenance   features  outperformed   Neighborhood  features   Supervised  Link  Predic0on   28  
29. •  Introduced  Friendship Similarity  and   Friendship Evenness  measures   •  Empirical  study  shows  most  users  prefer   to  maintain  different  friendships  in   different  OSNs,  while  keeping  only  a  small   clique  of  common  friends  across  OSNs   •  Reasonably  predict  links  in  Twiher  using   network  structural  properMes  from   Instagram  and  vice  versa.     Summary   29   Roy  LEE   Singapore  Management  University   roylee.2013@phdis.smu.edu.sg