Predicting Reciprocity in Social Networks

Presented at SocialCom 2011.

When looking at how people interact on Twitter, how can network factors help us predict which interactions are reciprocal (i.e. both parties participating), and which aren't (i.e. one user pestering another)? What factors are best in predicting reciprocity?

Transcript

1. Predicting
   in Social Networks
   Justin Cheng Daniel M. Romero
   Brendan Meeder Jon Kleinberg
   Reciprocity
   

   View Slide

2. View Slide

3. In real life, people engage
   in conversations
   

   View Slide

4. But lots of online
   communication is
   directed
   

   View Slide

5. An @-message is sent from
   one user to another
   Is this a conversation?
   

   View Slide

6. How about this?
   

   View Slide

7. Why is A contacting B?
   or
   @ladygaga  
   @random_fan  
   

   View Slide

8. Online relationships can be
   reciprocal or non-reciprocal
   

   View Slide

9. A superposition of two networks
   

   View Slide

10. Reciprocity can be
    subtle
    

    View Slide

11. Given characteristics of
    two users, can we
    determine whether they
    know each other?
    ?
     
    

    View Slide

12. How do we differentiate
    between symmetric and
    asymmetric interactions?
    ?
     
    

    View Slide

13. Can we predict if a
    relationship is
    reciprocal?
    ?
     
    

    View Slide

14. The @-message Graph
     
    ?
     
    

    View Slide

15. v w
    ?
     
    G
    Predicting symmetry
    (SYM)
    Given a graph
    and a node pair , predict
    both and
    exist
    or only one of these does
    {v, w}
    v ! w w ! v
    G
    

    View Slide

16. v w
    ?
     
    G
    Predicting a reverse
    edge (REV)
    Given the graph
    and that links to ,
    does link back to ?
    G
    v
    v
    w
    w
    

    View Slide

17. The edge
    is reciprocated both and have
    sent at least messages to each
    other
    (v, w)
    v w
    k
    The edge
    is unreciprocated if sent at
    least messages to
    but sent none in return
    (v, w)
    v
    w
    k
    w
    

    View Slide

18. sent messages  
    sent no messages  
    This relationship is reciprocated
    But this one is unreciprocated
    k
    sent messages  
    k
    sent messages  
    k
    

    View Slide

19. Identify reciprocated and
    unreciprocated edges in , and
    for each of these edges, try to
    predict whether the relationship is
    reciprocal.
    G
    v w
    G
    ?
     
    

    View Slide

20. Given the full network, hide only the
    link from to (if it exists).
    Try to predict whether the link
    actually exists.
    v w
    G
    ?
     
    v
    w
    

    View Slide

21. Outline
    Features that might predict reciprocity and
    how well they work
    – Individually,
    – Or in combination
    The structure of the reciprocated and
    unreciprocated sub-networks
    

    View Slide

22. Link reciprocity depends a lot on
    the relative status of two
    individuals
    @ladygaga  
    @average_joe  
    @average_jane  
    

    View Slide

23. Link reciprocity prediction
    vs.
    Link prediction
    Liben-Nowell and Kleinberg (2004)  
    ?
     
    

    View Slide

24. Link reciprocity prediction
    vs.
    Tie strength prediction
    Gilbert and Karahalios (2009)  
    S
     
    W
     
    

    View Slide

25. Link reciprocity prediction
    vs.
    Sign prediction
    Leskovec, Huttenlocher and Kleinberg (2010)  
    +
     
    –
     
    

    View Slide

26. What are good
    indicators of
    reciprocity?
    

    View Slide

27. For each feature, choose some
    threshold value above/below
    which we predict reciprocity to
    maximize accuracy.
    

    View Slide

28. Outdegree-indegree Ratio
    deg+(v)
    deg (v)
    /
    deg+(w)
    deg (w)
    v w
    c  
    c  
    deg (v)
    deg+(v) deg (w)
    deg+(w)
    

    View Slide

29. Individually,
    Outdegree-indegree ratio
    performed the best with
    82% accuracy
    

    View Slide

30. A smaller outdegree-indegree
    ratio indicated reciprocation
    deg+(v)
    deg (v)
    /
    deg+(w)
    deg (w)
    v w
    c  
    c  
    

    View Slide

31. A smaller outdegree-indegree
    ratio indicated reciprocation
    deg+(v) deg (w)
    deg (v) deg+(w)
    Ratio of Preferential Attachments  
    69% {  
    53% {  
    v w
    c  
    c  
    

    View Slide

32. Other features we tried
    •  Indegree and outdegree
    •  Incoming and outgoing messages
    •  Incoming message – indegree ratio (and out)
    •  Two-step paths in both directions
    •  Two-step paths ratio
    •  Mutual in-neighbors and out-neighbors
    •  Jaccard’s coefficient
    •  Adamic/Adar’s page similarity measure
    

    View Slide

33. Degree/Message
    Outdegree  
    Indegree  
    Outgoing Messages  
    Incoming Messages  
    And ratios between them  
    

    View Slide

34. Two-step Hops
    v  
    w  
    v  
    w  
    v  
    w  
    v  
    w  
    Mutual Neighbors  
    Two-step paths  
    

    View Slide

35. “Link prediction” features
    Jaccard’s coefficient =
    10 total neighbors  
    3 common neighbors  
    Common Neighbors
    Total Neighbors
    

    View Slide

36. “Link prediction” features
    Preferential attachment
    Product of indegree of and outdegree of
    v w
    v w
    c  
    

    View Slide

37. The Top 3
    Outdegree-indegree ratio  
    Two-step paths ratio  
    Indegree ratio  
    76%  
    76%  
    82%  
    

    View Slide

38. But the outdegree-indegree
    ratio and two-step paths ratio
    seem
    suspiciously similar…
    

    View Slide

39. v w
    c  
    c  
    Outdegree-indegree ratio  
    

    View Slide

40. v w
    Two-step paths ratio  
    

    View Slide

41. Marketer  
    Customers  
    Who’ll respond?  
    

    View Slide

42. It is better to know
    about than
    in predicting a reverse
    edge
    v
    w
    

    View Slide

43. So what happens when we use
    all the features we know?
    Link Pred Two-step
    Hops
    Deg/Msg Deg/Msg
    Ratio
    74% 80% 83% 86%
    

    View Slide

44. Decision Tree Accuracy on
    Sets of Features
    74%
    80% 83%
    86%
    v
    w
    v
    w
    

    View Slide

45. Decision Trees of Sets of
    Features
    80%
    74%
    83%
    86%
    

    View Slide

46. In a decision tree of all attributes,
    Outdegree-Indegree Ratio  
    86%
    accuracy  
    (STILL)  
    

    View Slide

47. Types of Edges
    Unreciprocated
    Reciprocated
    

    View Slide

48. Clustering Coefficient
    0.19
    0.02
    Reciprocated
    Unreciprocated
    

    View Slide

49. Are there two types of users
    on Twitter?
    “Reciprocators”  
    cf. informers and me-formers (Naaman et al.)  
    “Non-reciprocators”  
    

    View Slide

50. Types of Nodes
    65 30 5
    Both Reciprocated Edges Only Unreciprocated Edges Only
    

    View Slide

51. Most users take part in both
    reciprocated and
    unreciprocated interactions.
    @ladygaga  
    @average_joe  
    @friend_of_joe1  
    @friend_of_joe2  
    “I love your music @ladygaga!”  
    

    View Slide

52. Social, reciprocal relationships
    are associated with active,
    continued use of Twitter.
    

    View Slide

53. Features that approximate the
    relative status of two nodes
    seem most effective at
    predicting reciprocity between
    them.
    

    View Slide

54. Social networks are a superposition
    of reciprocated and unreciprocated
    relationships
    Reciprocity affects how we
    experience these sites
    Using network features, we can
    predict reciprocity in relationships
    

    View Slide

55. Thanks for Listening! Questions?
    Slide design heavily inspired by Paul Adams. Icons courtesy of The Noun Project.
    

    View Slide