decoupled smoothing on graphs
play

Decoupled smoothing on graphs Alex Chin, Yatong Chen , Kristen M. - PowerPoint PPT Presentation

Decoupled smoothing on graphs Alex Chin, Yatong Chen , Kristen M. Altenburger, Johan Ugander Stanford University The Web Conference, San Francisco May 17, 2019 1 Attribute prediction on graphs 2 Attribute prediction on graphs Given: Social


  1. Decoupled smoothing on graphs Alex Chin, Yatong Chen , Kristen M. Altenburger, Johan Ugander Stanford University The Web Conference, San Francisco May 17, 2019 1

  2. Attribute prediction on graphs 2

  3. Attribute prediction on graphs Given: Social Network ● 3 Yatong Chen, Stanford University, WWW 2019

  4. Attribute prediction on graphs Given: Social Network ● Male Female Labels for some subset nodes ● 4 Yatong Chen, Stanford University, WWW 2019

  5. Attribute prediction on graphs Given: Social Network ● Male ? Female Labels for some subset nodes ● ? Goal: ● Infer labels for unlabeled nodes ? 5 Yatong Chen, Stanford University, WWW 2019

  6. Semi-supervised learning Problem Given: Social Network ● Male ? Female Labels for some subset nodes ● ? Goal: ● Infer labels for unlabeled nodes ? 6 Yatong Chen, Stanford University, WWW 2019

  7. Approaches for attribute prediction ● Approach 1: Graph Smoothing based on Gaussian Random Field [Zhu, Ghahramani, Lafferty 2003] ○ Assumption : Gaussian Markov Random Field Prior on true label of all the nodes ○ Get the Bayes estimator of on unlabeled nodes under the GMRF prior ○ Be referred as ZGL later

  8. Approaches for attribute prediction ● Approach 2: LINK classification [Lu-Getoor 2003] Learn a function F: ○ F(row i in adjacency matrix) = i’s label Example F: regularized logistic regression ○ [0,1, ...0,1] Labeled [0,1, ...0,1] regularized logistic Labeled regression ? ? [0,0,..,0,1] 8 [1,0,...0,1] [1,0,...0,1] [0,0,..,0,1]

  9. ZGL’s assumption: Homophily ● Homophily [one-hop similarity]: ○ Individuals are similar to their friends 9 Yatong Chen, Stanford University, WWW 2019

  10. ZGL’s assumption: Homophily ● Homophily [one-hop similarity]: ○ ZGL assumes information of a given node decays smoothly across the topology of the graph (by imposing the GMRF prior) smooth! ? ? ? ? ? ? ? Labeled Labeled 10 10

  11. Homophily Assumption: NOT always necessary! ● [Altenburger-Ugander 2018] : ○ LINK does well even without assuming homophily ○ Homophily assumption is not necessary for inference to succeed ● ...but ZGL and a lot of other graph smoothing methods all assumes homophily. Can we do graph smoothing without it? ○ Yes (this talk) 11 Yatong Chen, Stanford University, WWW 2019

  12. A situation where Homophily assumption fails ● Gender example: Male Female 12 Yatong Chen, Stanford University, WWW 2019

  13. A situation where Homophily assumption fails ● Gender example: Want to predict the gender of the center node ○ Male Female ? 13 Yatong Chen, Stanford University, WWW 2019

  14. A situation where Homophily assumption fails ● Gender example: Want to predict the gender of the center node: ○ Assume Homophily (1-hop majority vote): false ■ Male Female ? ? 14 Yatong Chen, Stanford University, WWW 2019

  15. A situation where Homophily assumption fails ● Observation: there are difference between one’s identity and preference Identity: male Male Female Preference: female 15 Yatong Chen, Stanford University, WWW 2019

  16. Decoupled smoothing Method 16

  17. Decoupled smoothing: idea ● Idea: decoupling one’s “identity” and “preference” ● Use separate parameters to model them accordingly! Identity: male Male Female Preference: female Without homophily (or heterophily) assumption 17 Yatong Chen, Stanford University, WWW 2019

  18. Decoupled smoothing: idea ● Idea: decoupling one’s “identity” and “preference” 18 Yatong Chen, Stanford University, WWW 2019

  19. Decoupled smoothing: idea ● Idea: decoupling one’s “identity” and “preference” identity preference ● Intuition: a person’s identity will reveal information about their friend’s preference, and vice versa 19

  20. Decoupled smoothing: idea ● Idea: decoupling one’s “identity” and “preference” preference: can’t observe, identity how to reveal? ? ● Intuition: a person’s identity will reveal information about their friend’s preference, and vice versa 20

  21. Decoupled smoothing: model ● Intuition: a person’s identity will reveal information about their friend’s preference, and vice versa ● Assumption: 21 Yatong Chen, Stanford University, WWW 2019

  22. Decoupled smoothing: model ● Intuition: a person’s identity will reveal information about their friend’s preference, and vice versa ● Assumption: ● Goal: to obtain predictions for the identity ○ the preference is nuisance! ● Get the marginal prior for : ○ 22 Yatong Chen, Stanford University, WWW 2019

  23. Decoupled smoothing: impose a prior How to estimate W? 23 Yatong Chen, Stanford University, WWW 2019

  24. How to estimate W? ● Intuition: Node j’th preference will imply a 2-hop similarity between ○ node i and node k’s identities ? Node i Node j Node k 24

  25. How to estimate W? ● Intuition: Node j’th preference will imply a 2-hop similarity between ○ node i and node k’s identities Make use of the information of k when predicting i ○ ? Node i Node j Node k 25

  26. How to estimate W? ● Intuition: Node j’th preference will imply a 2-hop similarity between ○ node i and node k’s identities Make use of the information of k when predicting i ○ ● Assumption: ○ i’s 2-hop friend k has the distribution: ? Node i Node j Node k 26

  27. How to estimate W? ● Why as mean: Similarity among i and k ○ ● Why as variance ? ○ The more friends j have → the better its preference being revealed → the less uncertainty about the similarity between i and k Trust More! Node i Node j Node k 27

  28. How to estimate W? ● Assumption: 2-hop friend k has the distribution ○ Homogeneous standard error ○ ● Then W can be reduced to We get W! 28

  29. Decoupled smoothing: model ● Now we know everything about the marginal prior for : ○ ● Next step: ○ Compute the Bayes estimator of for unlabeled node and then make the prediction (recall ZGL) ● Done! 29 Yatong Chen, Stanford University, WWW 2019

  30. Relationship between Decoupled smoothing and some phenomenon/concept/method that are related to it 30

  31. Decoupled smoothing and Monophily ● The phenomenon of Monophily [Altenburger-Ugander 2018] ○ Two-hop similarity: individuals are similar to their friends’ friends ○ Innovative concept compared to Homophily Identity: male Male Preference: female Female 31

  32. Decoupled smoothing implies Monophily ● The phenomenon of Monophily [Altenburger-Ugander 2018] ○ Two-hop similarity: ■ similarity among the friends of a person is the result of personal preference ○ The 2 hop similarity phenomenon is implied by our decoupling smoothing idea! Identity: male Male Preference: female Female 32

  33. Decoupled smoothing and 2-hop MV ● Decoupled smoothing reduces to iterative 2-hop majority vote (under homogeneous standard error ): 2 hop Majority Vote (MV): Average over the labeled nodes in 2-hop friend sets 33

  34. Decoupled smoothing and ZGL Decoupled smoothing and ZGL ● ZGL: ● Decoupled smoothing: Assume Homophily Don’t assume Homophily ● ○ Prior: Prior: ● ○ Matrix A: Auxiliary matrix : ● ○ adjacency matrix Reduce to iteratively Reduce to iteratively ● ○ 1-hop majority vote 2-hop majority vote update method! update method! 34

  35. Empirical Results 35 Yatong Chen, Stanford University, WWW 2019

  36. Dataset ● Facebook 100: Single-day snapshots of Facebook in September 2005. ● Goal: gender prediction School Name Number of Nodes Number of Edges Amherst 2032 78733 Reed 962 18812 Haverford 1350 53904 Swarthmore 1517 53725 36

  37. Decoupled smoothing: empirical result Reed Swarthmore 37 Yatong Chen, Stanford University, WWW 2019

  38. Decoupled smoothing: empirical result ZGL ZGL Reed Swarthmore 38 Yatong Chen, Stanford University, WWW 2019

  39. Decoupled smoothing: empirical result ZGL ZGL Reed Swarthmore 39 Yatong Chen, Stanford University, WWW 2019

  40. Decoupled smoothing: empirical result ZGL ZGL Amherst Haverford 40 Yatong Chen, Stanford University, WWW 2019

  41. Why 2-hop Majority Vote beats Decoupled Smoothing? ZGL Amherst 41 Yatong Chen, Stanford University, WWW 2019

  42. Summary ● Introduce the idea of decoupling one’s “identity” and “preference” ● Justify/explain the phenomenon of 2-hop similarity without assuming homophily ● Open questions: ○ How to choose the weighted matrix W? ○ Why 2-hop Majority Vote outperforms decoupled smoothing: can you do better? 42 Yatong Chen, Stanford University, WWW 2019

  43. Questions? 43

  44. Thank you for your attention! 44

Recommend


More recommend