Published on

Published in: Technology
  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide
  • Excellent results on datasets from two popular blogs – Engadget and Huffington Post.
  • Research

    1. 1. Dynamics of Communication in Online Social Media<br />Munmun De Choudhury<br />PhD Student, Dept of Computer Sc & Engg, ASU<br />Advisor: Prof HariSundaram<br />
    2. 2. 2<br />Current State-of-the-Art<br />Face-to-face<br />Text Messaging<br />IP Telephony<br />Instant Messaging<br />Emails<br />
    3. 3. November 19, 2009<br />3<br />Modern communication modes<br />Slashdot<br />Facebook<br />Engadget<br />LiveJournal<br />Digg<br />Twitter<br />MetaFilter<br />Flickr<br />Reddit<br />Orkut<br />Blogger<br />YouTube<br />MySpace<br />
    4. 4. Some Social Media Statistics<br />4<br />
    5. 5. What are the impacts of such large-scale online social communication?<br /><ul><li>Microscopic level</li></ul>evolution of shared media characteristics (WWW 09)<br />information diffusion (WI 07, HT 08, SocialCom 09)<br /><ul><li>Macroscopic level</li></ul>group evolution (HT 08, CIKM 08, ICME 09)<br />network representations (Y! Research summer work)<br />5<br />
    6. 6. Motivating Applications …<br />What are the interesting conversations on the Blogosphere post Yahoo!’s Bing deal?<br />6<br />
    7. 7. Motivating Applications …<br />What has been the public buzz on the new Nikon D3000 SLR?<br />7<br />
    8. 8. Motivating Applications …<br />Which is the best news source to read about the recent Twitter crash?<br />8<br />
    9. 9. Problem 1<br />Collaborators:Ajita John, Doree Duncan Seligmann (Avaya Labs)<br />9<br />
    10. 10. Why do people repeatedly come back to the same You Tube video?<br />10<br />
    11. 11. They are returning to do more than watch the same video <br />11<br />
    12. 12. We think it is the conversations around the video they find interesting<br />12<br />
    13. 13. Themes and participants make conversations interesting<br />13<br />
    14. 14. An example of an interesting conversation …<br />14<br />14<br />
    15. 15. Conversational interestingness is not necessarily the popularity of a media object or preference of a topic<br />15<br />
    16. 16. Goal:<br />What causes a conversation to be interesting, that prompts a user to participate in the discussion on a posted video?<br />Our Contributions<br /><ul><li>Approach:
    17. 17. Detect conversational themes.
    18. 18. Determine interestingness of participants and interestingness of conversations based on a random walk model.
    19. 19. Measure the consequence of a conversations.
    20. 20. Excellent results on a dataset from YouTube.</li></ul>16<br />
    21. 21. How do we determine interestingness?<br />17<br />
    22. 22. 18<br />Conversational Themes<br />Conversational themes are sets of salient topics associated with conversations at different points in time. <br />bag of words (chunks, i,t) over time slices<br />……<br />t1<br />t2<br />tQ<br />……<br />Theme models, θj<br />= ?<br />conversation i<br />
    23. 23. Theme Model<br />19<br /><ul><li>Temporal Regularization
    24. 24. A word w in the chunk can be attributed either to the textual context λi,t, or the time slice t
    25. 25. Smoothness of theme models over time</li></ul>Chunk context<br />Temporal Smoothing<br />Time context<br />Theme distribution<br />Theme strength<br />time<br />time<br />
    26. 26. Theme Model (Contd.)<br />Co-participation based Regularization<br />If several participants comment on a pair of chunks, their theme distributions are likely to be close to each other.<br />conversation ci, f(θj|ci)<br />ωi,m<br />conversation cm, f(θj|cm)<br />20<br />
    27. 27. Interestingness of Conversations<br />21<br />Impact due to<br />participants<br />Impact due to<br />themes<br />ψ<br />1ψ<br />Relationship between participants and conversations<br />X<br />ψ<br />Interestingness of participants<br />Interestingness of conversations<br />Conversational Theme Distribution<br />X<br />1─ψ<br />Theme Strength<br />X<br />Interestingness of conversations<br />t<br />t─1<br />
    28. 28. Interestingness of Participants<br />22<br />1β<br />β<br />Past communication<br />Preference<br />
    29. 29. Interestingness of participants and conversations mutually reinforce each other<br />23<br />
    30. 30. 24<br />Joint Optimization of Interestingness<br />A joint optimization framework, which maximizes the two interestingness measures for optimal X=(α1, α2, α3, ) and also incorporates temporal smoothness: <br />Interestingness of participants<br />Interestingness of conversations<br />Regularization of participants’ interestingness<br />Regularization of conversations’ interestingness<br />
    31. 31. What happens after a conversation becomes interesting?<br />25<br />
    32. 32. 26<br />activity<br />t<br />t+δ<br />cohesiveness<br />participant<br />Participant <br />activity<br />conversation<br />t<br />t+δ<br />Participant <br />Cohesiveness<br />t<br />t+δ<br />Thematic <br />interestingness<br />Three consequence metrics of interestingness <br />
    33. 33. 27<br />YouTube Dataset<br />‘News & Politics’ category on YouTube – rich communication on highly dynamic events.<br />132,348 videos<br />~ 9M unique participants<br />~ 89M comments<br />15 weeks from June 20, 2008 to September 26, 2008<br />
    34. 34. November 19, 2009<br />28<br />
    35. 35. 29<br />Analysis of Interestingness of Participants<br />Interestingness of participants is less affected by number of comments during significant external events <br />
    36. 36. 30<br />Analysis of Interestingness of Conversations<br /><ul><li>Mean interestingness of conversations increases during periods of several external events; however, certain highly interesting conversations always occur at different weeks irrespective of events.</li></li></ul><li>Evaluation using Consequences<br />Interestingness is computed using five techniques – <br />our method with temporal smoothing (I1), <br />our method without temporal smoothing (I2) and <br />the three baseline methods, <br />B1 (comment frequency), <br />B2 (novelty of participation), <br />B3 (co-participation based PageRank).<br />31<br />31<br />
    37. 37. Conclusions<br />cohesiveness<br />participant<br />t<br />t+δ<br />Participant Cohesiveness<br />Summary <br />Why do people repeatedly come back to the same YouTube video?<br />Our method can explain future consequences<br />Future Work<br />User subjectivity<br />Personalized recommendations<br />
    38. 38. Problem 2<br />Collaborators:Ajita John, Doree Duncan Seligmann (Avaya Labs)<br />
    39. 39. Social Synchrony<br />Goal:<br />a framework for predicting social synchrony in online social media over a period of time into the future. <br />Approach:<br />Operational definition of social synchrony.<br />Learning – a dynamic Bayesian representation of user actions based on latent states and contextual variables.<br />Evolution – evolve the social network size and the user models over a set of future time slices to predict social synchrony.<br />Excellent results on a large dataset from the popular news-sharing social media Digg. <br />
    40. 40. Observations from Digg<br />Topic ‘Olympics’ is observed to exhibit synchrony where old users continue to be involved in the action of digging stories, as well as large number of new users join in the course of time (Sept 3-Sept 13). <br />
    41. 41. 36<br />Experiments on Digg Data<br />Digg dataset<br />August, September 2008 <br />21,919 users, 187,277 stories, 7,622,678 diggs, 687,616 comments and 477,320 replies.<br />Six sample topics – four inherently observed to have synchrony.<br />
    42. 42. Problem 3<br />Collaborators:Ajita John, Doree Duncan Seligmann (Avaya Labs)<br />
    43. 43. Individuals<br />Groups<br />Prototypical groups<br />How do groups evolve centered around communication in the Blogosphere?<br />38<br />
    44. 44. Goal:<br />How do we characterize communication at the level of individuals and groups?<br />How do we extract groups?<br />Which are the prototypical groups given a topic?<br />39<br />Mining Prototypical Social Groups<br /><ul><li>Approach:
    45. 45. Extract individual characteristics based on communication.
    46. 46. Extract groups based on a random walk model based clustering algorithm.
    47. 47. Determine prototypical groups – composition entropy, topic divergence.</li></ul>39<br />
    48. 48. 40<br />
    49. 49. Evaluation<br />B1 B2 SVR Actual<br />B1 B2 SVR Actual<br />B1 B2 SVR Actual<br />B1 B2 SVR Actual<br />Google getting into e-books (Jan 22)<br />Microsoft buying Yahoo rumor (May 4)<br />Apple Beatles settles trademark lawsuit (Feb 5)<br />Nokia E61i, E65 (Jun 25)<br />Google outbids Microsoft for Dell bundling deal (May 25)<br />Nokia recalls 46M batteries (Aug 14)<br />Open Text Corporation extends alliance with Microsoft (Aug 20)<br />iPhone release (Jun 29)<br />The FTC to conduct antitrust review of Google and DoubleClick deal (May 29)<br />Nokia NGage gaming service (Aug 29)<br />Microsoft To Acquire Parlano (Aug 30)<br />Ipod classic, ipod touch (Sep 5)<br />Microsoft loses EU anti-trust appeal (Sep 17)<br />US STOCKS-Tech drag trips Dow, Nasdaq extends drop (Nov 29)<br />Microsoft Xbox 360 Arcade release (Oct 23)<br />APPLE<br />GOOGLE<br />MICROSOFT<br />NOKIA<br />Visualization of stock movements with time on vertical scale. B1 and B2 are two baseline techniques. Blue bubbles indicate positive movement and red bubbles negative movements. Sizes of the bubbles represent magnitude of movement. The SVR prediction is found to follow the movement trend very closely with an error of 26.41 %.<br />
    50. 50. Problem 4<br />Collaborators:Winter Mason, Jake Hofman, Duncan Watts (Yahoo! Research)<br />
    51. 51. How do we infer optimal network structures based on social communication?<br />43<br />
    52. 52. Empirical Observations<br />
    53. 53. Prediction Tasks<br />45<br />
    54. 54. Summary…<br />46<br />