Measuring User Influence, Susceptibility and Cynicalness in Sentiment Diffusion

Diffusion in social networks is an important research topic lately due to massive amount of information shared on social media and Web. As information diffuses, users express sentiments which can affect the sentiments of others. In this paper, we analyze how users reinforce or modify sentiment of one another based on a set of inter-dependent latent user factors as they are engaged in some diffusion of event information. We introduce these sentiment-based latent factors, namely influence, sus- ceptibility and cynicalness. We also propose the ISC model to relate the three factors together and develop an iterative computation approach to derive them simultaneously. We evaluate the ISC model by conducting experiments on two separate sets of Twitter data collected from two real world events. The experiments show the top influential users tend to stay consistently influential while susceptibility and cynicalness of users could change significantly across events.

Transcript

1. Measuring  User  Influence,  Suscep2bility  and   Cynicalness  in  Sen2ment  Diffusion   Presented  by:  Roy  LEE   Roy  Ka-­‐Wei  LEE  &  Ee-­‐Peng  LIM  
2. Outline   •  Introduc8on   o  Diffusing  Sen8ments  in  Social  Networks   o  Research  Objec8ves   o  Real-­‐World  Applica8on   •  Modeling  Sen8ment  Diffusion   o  Sen8ment  Diffusion   o  ISC  Model   o  Model  Computa8on   •  Experiment   o  Dataset   o  Results   •  Conclusion   o  Contribu8on   o  Future  Works   2  
3. 3   INTRODUCTION  
4. Diffusing  Sen2ments   • Emo8on  induces  and  boost  social   transmission  of  informa8on1     • Similar  observa8ons  can  be  made  for   informa8on  diffusion  in  social  networks   4   1K.  Peters  et  al.  Talking  about  others:  Emo4onality  and  the  dissemina4on  of  social  informa4on.  EJSP,  2009.  
5. Diffusing  Sen2ments   5   Time  
6. Research  Objec2ves   • Iden8fying  latent  user  aDributes  that   contribute  to  sen8ment  diffusion  in  social   network   o Influence   o Suscep2bility   o Cynicalness   •  Modeling  the  inter-­‐dependency  among  the  latent   user  aWributes   6  
7. Real-­‐World  Applica2ons   • Sensing  Social  Events   o Detect  change  in  user’s  influence,  suscep8bility  and   cynicalness  towards  certain  events  content   o Example:  Elec8ons,  radical  terrorism  ideologies     • Marke8ng     o Influen8al  users  to  swing  sen8ments  in  favor  of   products  and  brands   o Target  suscep8ble  users  and  avoid  cynical    ones  to  gain   posi8ve  sen8ments  on  marke8ng  campaign     7  
8. 8   MODELING  SENTIMENT  DIFFUSION  
9. Sen2ment  Diffusion   • Diffusion  via  retweet  is  too  restric8ve  for   sen8ment  diffusion       • Event-­‐based  sen8ment  diffusion   o Collect  the  relevant  tweets  for  an  event  (e.g.  hashtag)   o Study  the  diffusion  of  relevant  tweets  among  users   9  
10. •  User  u  diffused  tweet  sen8ment  x  to  follower  v  when:   1.  u  adopts  x  before  v  adopts  the  same  or  opposing    tweet   sen8ment  x’   2.  v  is  a  follower  of  u  when  v  adopts  x’   3.  u’s  tweet  with  sen8ment  x  is  the  latest  received  tweet  on   v’s  TwiWer  8meline   Sen2ment  Diffusion  Example   10   Types  of  sen2ment   +:  posi8ve  tweet  sen8ment   0:  neutral/no  tweet  sen8ment   -­‐:  nega8ve  tweet  sen8ment   8me  t   8me  t’  
11. • Influence-­‐Suscep8bility-­‐Cynicalness  (ISC)  Model   o Measures  user  influence,  suscep8bility  and  cynicalness   simultaneously   • Three  inter-­‐dependent  modelling  principles   1.  Influen2al  user  can  get  non-­‐suscep8ble  and  cynical  users  to   change  and  adopt  the  same  tweet  sen8ment  diffused  by  him   2.  Suscep2ble  user  adopts  the  same  tweet  sen8ment  diffused   by  non-­‐influen8al  user   3.  Cynical  user  adopts  opposite  tweet  sen8ment  diffused  to  him   by  non-­‐influen8al  user   ISC  Model   11  
12. Tweet  level  User-­‐User  Influence   12   • Influence  of  user  u  is  measured  by   o The  sen8ment  change  u  has  cause  follower  v  when  u   diffuse  a  tweet  sen8ment  to  v  
13. Tweet level User-User Influence   13     X(u)   X(v)   1(+ve)   0(neutral)   -­‐1(-­‐ve)   1(+ve)   0   1  if  x’(v)=1   2  if  x’(v)=1;  1  if  x’(v)=0   0(neutral)   0.5  if  x’(v)=0   0   0.5  if  x’(v)=0   -­‐1(-­‐ve)   2  if  x’(v)=-­‐1;  1  if  x’(v)=0   1  if  x’(v)=-­‐1   0   •  fs (),  returns  the  sen8ment  change  when  v  adopts  same  sen8ment  as  u  
14. Tweet level User-User Influence   14     X(u)   X(v)   1(+ve)   0(neutral)   -­‐1(-­‐ve)   1(+ve)   0   1  if  x’(v)=1   2  if  x’(v)=1;  1  if  x’(v)=0   0(neutral)   0.5  if  x’(v)=0   0   0.5  if  x’(v)=0   -­‐1(-­‐ve)   2  if  x’(v)=-­‐1;  1  if  x’(v)=0   1  if  x’(v)=-­‐1   0   •  fs (),  returns  the  sen8ment  change  when  v  adopts  same  sen8ment  as  u  
15. Tweet level User-User Influence   15     X(u)   X(v)   1(+ve)   0(neutral)   -­‐1(-­‐ve)   1(+ve)   0   1  if  x’(v)=1   2  if  x’(v)=1;  1  if  x’(v)=0   0(neutral)   0.5  if  x’(v)=0   0   0.5  if  x’(v)=0   -­‐1(-­‐ve)   2  if  x’(v)=-­‐1;  1  if  x’(v)=0   1  if  x’(v)=-­‐1   0   •  fs (),  returns  the  sen8ment  change  when  v  adopts  same  sen8ment  as  u  
16. Tweet level User-User Influence   16     X(u)   X(v)   1(+ve)   0(neutral)   -­‐1(-­‐ve)   1(+ve)   0   1  if  x’(v)=1   2  if  x’(v)=1;  1  if  x’(v)=0   0(neutral)   0.5  if  x’(v)=0   0   0.5  if  x’(v)=0   -­‐1(-­‐ve)   2  if  x’(v)=-­‐1;  1  if  x’(v)=0   1  if  x’(v)=-­‐1   0   •  fs (),  returns  the  sen8ment  change  when  v  adopts  same  sen8ment  as  u  
17. Tweet level User-User Influence   17     X(u)   X(v)   1(+ve)   0(neutral)   -­‐1(-­‐ve)   1(+ve)   0   1  if  x’(v)=1   2  if  x’(v)=1;  1  if  x’(v)=0   0(neutral)   0.5  if  x’(v)=0   0   0.5  if  x’(v)=0   -­‐1(-­‐ve)   2  if  x’(v)=-­‐1;  1  if  x’(v)=0   1  if  x’(v)=-­‐1   0   •  fs (),  returns  the  sen8ment  change  when  v  adopts  same  sen8ment  as  u  
18. Tweet level User-User Influence   18     X(u)   X(v)   1(+ve)   0(neutral)   -­‐1(-­‐ve)   1(+ve)   0   1  if  x’(v)=1   2  if  x’(v)=1;  1  if  x’(v)=0   0(neutral)   0.5  if  x’(v)=0   0   0.5  if  x’(v)=0   -­‐1(-­‐ve)   2  if  x’(v)=-­‐1;  1  if  x’(v)=0   1  if  x’(v)=-­‐1   0   •  fs (),  returns  the  sen8ment  change  when  v  adopts  same  sen8ment  as  u   •  fo (),  returns  the  sen8ment  change  when  v  adopts  opposite  sen8ment  as  u     X(u)   X(v)   1(+ve)   0(neutral)   -­‐1(-­‐ve)   1(+ve)   2  if  x’(v)=-­‐11;  1  if  x’(v)=0   1  if  x’(v)=0   0   0(neutral)   0   0   0   -­‐1(-­‐ve)   0   1  if  x’(v)=1   2  if  x’(v)=1;  1  if  x’(v)=0  
19. Aggregated  Tweet  Level  User  Influence   19   • Influence  of  user  u  is  measured  by   o The  sen8ment  change  u  has  cause  follower  v   o Average  sen8ment  change  for  all  followers  who  u  has   diffused  a  tweet  sen8ment  to  
20. Aggregated  User  Influence   20   • Influence  of  user  u  is  measured  by   o The  sen8ment  change  u  has  cause  follower  v   o Average  sen8ment  change  for  all  followers  who  u  has   diffused  a  tweet  sen8ment  to   o Sum  of  the  average  sen8ment  changes  for  all  tweet   sen8ment  diffused  
21. User  Influence  in  ISC  Model   21   • Influence  of  user  u  is  measured  by   o The  sen8ment  change  u  has  cause  follower  v   o Average  sen8ment  change  for  all  followers  who  u  has   diffused  a  tweet  sen8ment  to   o Sum  of  all  the  average  sen8ment  changes  for  all   followers  who  u  has  diffused  tweet  sen8ments  to   o Take  into  considera8on  the  suscep8bility  and   cynicalness  of  followers  
22. ISC  Model   22  
23. ISC  Model  Computa2on   23   • Simultaneously  and  Itera8vely  compute  I(u),   S(v)  and  C(v)  
24. 24   EXPERIMENT  
25. Experiment  Overview   1.  Collect  event   related  tweets  by   relevant  hashtags   and  keywords   2.  Classify  sen2ments   of  Tweets  using   Sen2ment1401   classifier     3.  Apply  ISC   Model  to  data   and  compute   users’  Influence,   Suscep2bility  and   Cynicalness  score   1hWp://help.sen8ment140.com/api   25  
26. Dataset   26   5,570     Users   LiDle  India  Riot   Singapore  Haze   16,190  Tweets   18,933  Tweets   Hashtag:  #riot,  #liWleindiariot  etc   Hashtag:  #sghaze,  #haze  etc   •  A  riot  that  occurred  in   Singapore  December  2013   •  The  Singapore  haze  crisis   occurred  between  June  to   July  2013  
27. Log-­‐Log  ISC  Score  Distribu2on   27   Riot   Haze  
28. •  Jaccard  similarity  between  top  k%  users  in  Haze   and  Riot  events       Comparison  of  Haze  and  Riot  ISC  Score   28  
29. •  Comparing  Influence  measure  in  ISC  model  with  In-­‐ Degree  and  PageRank  measures  using  Pearson   Correla8on     Comparison  Influence  Measures   29  
30. Top  10  Influen2al  Users   30   Rank Username Followers 1   @livinginsg   21446   2   @TheDeeKosh   19559   3   @SoSingaporean   129142   4   @asonofapeach   7102   5   @TODAYonline   119184   6   @mrbrown   99607   7   @toshrock   66504   8   @RidhwanAzman   18948   9   @RovinNa8on   6370   10   @ChannelNewsAsia   172418   Overlap  top  10  users  in  both  Riot  and  Haze  data   Rank Username Followers 1   @SchoolprobIems   57606   2   @NaomiNeo_   69884   3   @cabbitowl   8514   4   @RidhwanAzman   18948   5   @livinginsg   21446   6   @speishi   33019   7   @SlickScribe   24849   8   @flyirene   19312   9   @TODAYonline   119184   10   @fakeMOE   22064   Riot   Haze  
31. 31   CONCLUSION  
32. • Introduced  user  influence,  suscep8bility  and   cynicalness  as  latent  user  aWributes  affec8ng   sen8ment  diffusion   • Proposed  a  novel  model  (ISC)  to  model  the   inter-­‐dependency  between  the  user   aWributes  and  measure  them  simultaneously   Research  Contribu2on   32  
33. • Consider  the  intrinsic  interest  of  users   • Study  more  detailed  emo8ons  such  as  fear,   anger  etc   • Extend  study  to  mul8ple  social  networks   informa8on  diffusion   Future  Works   33  
34. 34   Thank  You   Roy  LEE   PhD  Student  @  School  of  Informa4on  Systems   Singapore  Management  University   Contact:   roylee.2013@phdis.smu.edu.sg