can who edits what predict edit survival
play

Can Who-Edits-What Predict Edit Survival? Batuhan Yardm, Victor - PowerPoint PPT Presentation

Can Who-Edits-What Predict Edit Survival? Batuhan Yardm, Victor Kristof , Lucas Maystre, Matthias Grossglauser I nformation and N etwork Dy namics Lab (indy.ep fl .ch) August 23, 2018 KDD18 London Peer-production systems Emergence of


  1. Can Who-Edits-What Predict Edit Survival? Batuhan Yardım, Victor Kristof , Lucas Maystre, Matthias Grossglauser I nformation and N etwork Dy namics Lab (indy.ep fl .ch) — August 23, 2018 — KDD18 – London

  2. Peer-production systems Emergence of self-organizing, crowd-sourced projects online. Distributed vs. centralized production. � 2

  3. Problem Projects are victims of their own success : problems arise with increasing scale . Alan Turing « Alan Turing was an English « Blah blih bluh!@!? » computer scientist… » ??? ??? ??? Predict quality of contributions. Help project maintainers in their work. Help users match their interests. � 6

  4. Typical approaches User reputation systems Highly specialized predictors INTERANK 42 58 23 #words timestamp user IP Simple Simple Complex General General Specialized Not accurate Accurate Accurate � 7

  5. Model: INTERANK Experiment: Wikipedia Experiment: Linux

  6. Model: INTERANK Experiment: Wikipedia Experiment: Linux

  7. INTERANK: basic variant Model the probability p ui that an edit made by user u on item i is successful … …as a game between user u and item i (inspired by Bradley-Terry models). 1 1 + exp[ − ( s u − d i + b )], s u , d i , b ∈ R p ui = If s u increases , p ui increases . If d i increases , p ui decreases . Skill of user u Di ffi culty of item i Bias Informally: • Skill quanti fi es ability of user to make a contribution . • Di ffi culty quanti fi es how « resistant » to contributions a particular item is. � 10

  8. INTERANK: full variant Too simplistic: if user u is more skilled than user v , then p ui > p vi for all items i . Need to capture the interactions between users and items. 1 x u , y i ∈ R D p ui = u y i + b )], 1 + exp[ − ( s u − d i + x ⊺ If and are close , p ui increases . x u y i Embedding of Dimension of Embedding of user u latent space item i Informally: • describes the set of skills displayed by user u . x u • describes the set of skills needed to edit item i . y i � 11

  9. INTERANK: learning A dataset of K observations consists of triplets ( u k , i k , q k ) , k =1 , …, K . 𝒠 The outcome q k {0, 1} encodes whether an edit by user u on item i survives. ∈ ∑ [ − q log p ui − (1 − q )log(1 − p ui ) ] − ℓ ( θ ; 𝒠 ) = ( u , i , q ) ∈𝒠 basic: full: θ = [ s 1 , . . . , s N , d 1 , . . . , d M ] θ = [ s 1 , . . . , s N , d 1 , . . . , d M , { x u 1 , . . . , x uD } N u =1 , { y i 1 , . . . , y iD } M i =1 ] basic: log-likelihood is convex full: bilinear term breaks convexity In practice: • We do not observe any convergence issues. • We reliably fi nd good model parameters using Stochastic Gradient Descent . � 12

  10. Model: INTERANK Experiment: Wikipedia Experiment: Linux

  11. Wikipedia Edition # users # articles # edits French 5.5M 1.9M 65M Turkish 1.4M 0.3M 8.8M Competing approaches Average: GLAD: [Whitehill et al., 2009] 1 # good edits 1 p ui = p = 1 + exp[ − ( s u / d i + b )] # total edits 1 + exp[ − ( s u + b )] INTERANK Naive predictor User-only: [Adler & de Alfaro, 2007] ORES: [Halfaker & Taraborelli, 2015]: Uses over 80 content-based and system- 1 p u = based features. Di ff erent for Turkish and 1 + exp[ − ( s u + b )] French. Specialized predictor Reputation system � 14

  12. ���� �������� ������� ��� ��� ��� ��� ��� ���� �������� ���� ��������� �������� ���� ������ �������� ����� ������ ��������� ������ ��� ��� ��� ��� ��� ��� ��� ������� ��������� ��� ���� ��������� ������� �������� ���� �������� ����� ������ ��������� ������ ��� ��� ��� ��� ��� ��� ������� ��������� ��������� ��� ��� ��� ��� ��� ��������� ������ ��� ��� ��� ��� ��� ��� ������� ��������� ��������� ��� ��� ��� ��� ��� ��� ������ ��� ��� ��� ��� ������ ��������� ��� ��� ��� ��� ��� ��� ��� ������ ��� ��� ��� �������� ����� ��� ��� ��� ���� �������� ���� ��������� ������� �������� ���� ��� Wikipedia: results ORES has the best AUPRC and INTERANK full has the best log-likelihood . � 15

  13. Wikipedia: difficulty parameter d i Compare: Rank Title Percentile of di d i 1 Ségolène Royal 99.840 % • Manual ranking of controversial articles 2 Unidenti fi ed fl ying object 99.229 % [Yasseri et al., 2014] 3 Jehovah’s Witnesses 99.709 % 4 Jesus 99.953 % • Ranking of di ffi culty parameter d i as 5 Sigmund Freud 97.841 % learned by INTERANK 6 September 11 attacks 99.681 % 7 Muhammad al-Durrah incident 99.806 % 8 Islamophobia 99.787 % 9 God in Christianity 99.712 % 10 Nuclear power debate 99.304 % � 16

  14. ��� ���� ��� ��� ��� ��� ��� ������ ������ ��������� �������� ����� �������� ���� ������� ��������� ���� �������� ������� ��������� ������� ��������� ��������� ��� ��� ��� ��� ��� ��� ������ ��� ��� ��� ��� ��������� ��� ��� ���� �������� ���� ��������� ������� �������� ���� �������� ����� ������ ��������� ������ ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ��� ������ ��� ��� ��� ��� ��� ��� Wikipedia: latent factors Y = [ y i ] Highest Lowest Seven Wonders of the World Harry Potter’s magic list Thomas Edison List of programs broadcasted by Star TV s e l s c e Cell Bursaspor 2011-12 season i t l r c Justine Henin a i t r e Julie Halard a Mustafa Kemal Atat ü rk Kral Pop TV Top 20 r u e Virginia Wade r t l u William Shakespeare u Marcelo Melo t Albert Einstein Death Eater c l M. de Robespierre u r … c a Nelson Mandela l h u Democracy Heroes (TV series) g p Charlemagne i H o P … Isaac Newton List of programs broadcasted by TV8 Mehmed the Conqueror Karadayı TV & teen culture French municipality Leonardo da Vinci Show TV Tennis-related Other Louis Pasteur List of episodes of Kurtlar Vadisi Pusu � 17

  15. Model: INTERANK Experiment: Wikipedia Experiment: Linux

  16. ������� ��� ������ ������ �������� ���� �������� ����� ����� ������ ��������� ��� ��� ��� ��� ���� ��� ������ ��� ��� ��� ��� ��� ��� ��������� Linux # developers # subsystems # patches % accepted 9 672 394 619 419 34.12 % Dataset from [Jiang et al., 2013]. Developers submit patches to subsystems. A patch is accepted if it makes it into a Linux release. Specialized classi fi er: random forest using 21 features. � 19

  17. Linux: difficulty parameter Di ffi culty Subsystem % accepted +2.66 1.9 % usr Avg. number of commits in +1.33 7.8 % include Core components last quartile = 833 +1.04 16.0 % lib +1.01 34.3 % drivers/clk +0.87 17.7 % include/trace -0.80 45.4 % arch/mn10300 Avg. number of commits in -0.94 73.0 % net/nfc Peripheral components fi rst quartile = 687 -0.99 44.3 % drivers/ps3 -1.08 43.1 % net/tipc -1.19 78.3 % drivers/addi-data « Higher number of commits leads to lower acceptance rate. » [Jiang et al., 2013] � 20

  18. Conclusion INTERANK provides a new point in the solution space . Specialized predictors INTERANK Accuracy Reputation systems Generality Easy to implement and computationally inexpensive. Yields insights into collaborative projects. � 21

  19. Can who-edits-what predict edit survival? YES!

  20. Thank you! /lca4/interank

Recommend


More recommend