The Web Conference 2020 - Participation Report

Wantedly の Data 系メンバー 5 名で WWW 2020 に参加しました！
中でもソーシャルネットワークにおけるつながりの活用に関する論文をいくつか参加報告会で紹介しました。
https://connpass.com/event/174856/

Transcript

1. ©2020 Wantedly, Inc.
   Social Networks ʹ͓͚Δ
   influence ͱ fake accounts ʹؔ͢Δ࿦จ঺հ
   The Web Conference 2020 ࢀՃใࠂձ
   Apr 30, 2020 - Naomichi Agata
   

   View Slide

2. ©2018 Wantedly, Inc.
   Naomichi Agata
   - 2018/04~ Wantedly, Inc.
   - Wantedly People / Data Chapter Leader / Engineer
   - σʔλΛ࢖ͬͯαʔϏεΛվળ͢ΔνʔϜ
   - Wantedly Platform ্ͷʮͭͳ͕ΓʯΛվળ͢Δɾ׆༻͢Δ͜ͱ͕࠷ۙͷϛογϣϯ
   - kaggle expert
   @agatan_
   ࣗݾ঺հ
   

   View Slide

3. ©2020 Wantedly, Inc.
   ঺հ͢Δ࿦จ
   • Social Network Influence Ranking via Embedding Network Interactions
   for User Recommendation
   •
   Hongbo Bo, Ruan McConville, Jun Hong, Weiru Liu, 2020.
   •
   https://dl.acm.org/doi/abs/10.1145/3366424.3383299
   • On Twitter Purge: A Retrospective Analysis of Suspended Users
   •
   Farhan Asif Chowdhury , Lawrence Allen , Mohammad Yousuf , Abdullah Mueen, 2020.
   •
   https://dl.acm.org/doi/abs/10.1145/3366424.3383298
   • Friend or Faux: Graph-Based Early Detection of Fake Accounts
   on Social Networks
   • Adam Breuer, Roee Eilat, Udi Weinsberg, 2020.
   • https://dl.acm.org/doi/abs/10.1145/3366423.3380204
   

   View Slide

4. ©2020 Wantedly, Inc.
   Social Network Influence Ranking
   via Embedding Network Interactions
   for User Recommendation
   ঺հ࿦จຊ໨
   Hongbo Bo, Ruan McConville, Jun Hong, Weiru Liu
   

   View Slide

5. ©2020 Wantedly, Inc.
   ໨తSNS্ͷӨڹྗͷఆྔԽʹϢʔβؒͷinteractionΛߟྀ͍ͨ͠
   ঺հ࿦จ (1): Social Network Influence Ranking via Embedding Network Interactions for User Recommendation
   • Өڹྗͷߴ͞ (Influence) Λద੾ʹఆྔԽͯ͠ਪનʹ໾ཱ͍ͯͨ
   • Ծఆ: Ϣʔβ͸Өڹྗͷߴ͍ϢʔβΛϑΥϩʔ͠΍͍͢
   • ϑΥϩʔɾϑΥϩϫʔͷؔ܎͚ͩͰ͸ෆे෼
   

   View Slide

6. ©2020 Wantedly, Inc.
   ఏҊख๏ͷ֓ཁ
   • άϥϑ্ͷϊʔυͷॏཁ౓Λܾఆ͢ΔΞϧΰϦζϜͰ͋Δ PageRank Λ
   ϕʔεͱ͢Δ EIRank ΛఏҊ
   • follow ؔ܎ͻͱͭͻͱͭʹॏΈΛ͚ͭΔ
   • interaction ৘ใΛ༻͍ͯɼΑΓີ઀ͳ follow ؔ܎ʹ͸ߴ͍ॏΈΛׂΓ౰ͯΔ (= ॏࢹ͢Δ)
   • interaction graph ্ͷ node2vec Λར༻͠ɼfollow ؔ܎ͷڧ͞ΛఆྔԽ
   ঺հ࿦จ (1): Social Network Influence Ranking via Embedding Network Interactions for User Recommendation
   

   View Slide

7. ©2020 Wantedly, Inc.
   લఏ: PageRank
   • άϥϑશମʹରͯ͠ɼ֤ϊʔυ͕Ͳͷ͘Β͍ॏཁ͔Λఆྔతʹද͢
   • ͨ͘͞ΜϑΥϩʔ͞Ε͍ͯΔͱߴ͘ͳΔ
   • ॏཁ౓͕ߴ͍ਓʹϑΥϩʔ͞Ε͍ͯΔͱߴ͘ͳΔ
   • ͨͩ͠ϑΥϩʔ਺͕ଟ͍ਓʹϑΥϩʔ͞Εͯ΋্ঢ෯͸খ͍͞
   ঺հ࿦จ (1): Social Network Influence Ranking via Embedding Network Interactions for User Recommendation
   

   View Slide

8. ©2020 Wantedly, Inc.
   EIRank
   • interaction ΛΤοδͱΈͳͨ͠ॏΈ෇͖άϥϑΛߏங͠ɼ
   ͦͷάϥϑ্ͰͷϢʔβؒͷڑ཭ (≒ ਌ີ౓) Λ node2vec ͰఆྔԽ͢Δ
   ঺հ࿦จ (1): Social Network Influence Ranking via Embedding Network Interactions for User Recommendation
   

   View Slide

9. ©2020 Wantedly, Inc.
   EIRank
   • node2vec ͷڑ཭Λ΋ͱʹϑΥϩʔ͝ͱͷ਌ີ౓Λܭࢉ
   • PageRank Ͱ͸Ұ༷ʹ 1/N ͱ͍ͯͨ͠෦෼Λ਌ີ౓ʹஔ͖׵͑Δ
   ঺հ࿦จ (1): Social Network Influence Ranking via Embedding Network Interactions for User Recommendation
   

   View Slide

10. ©2020 Wantedly, Inc.
    ࣮ݧɾධՁ
    • Twitter ͷϑΥϩʔਪનͰධՁ͍ͯ͠Δ
    • ਖ਼ྫ 1, ෛྫ 10 ͷ 11 ਓΛฒͼସ͑ͯɼਖ਼ྫͷฏۉॱҐΛൺֱ
    • Influence ͷӨڹ͕େ͖͍ Platform Ͱ͋Δ͜ͱ͕Α͘Θ͔Δ…
    ঺հ࿦จ (1): Social Network Influence Ranking via Embedding Network Interactions for User Recommendation
    

    View Slide

11. ©2020 Wantedly, Inc.
    ·ͱΊ
    • Influence ͷఆྔԽʹϢʔβؒͷ interaction ΛՃຯ͢Δ EIRank ΛఏҊ
    • Ϣʔβਪન΁ͷԠ༻ΛධՁ͠ɼطଘख๏Λ௒͑ΔੑೳΛ֬ೝͨ͠
    ײ૝
    • ͭͳ͕Γ͕ฏ౳Ͱͳ͍͜ͱΛఆྔԽ͍͓ͯͯ͠΋͠Ζ͔ͬͨ
    • ෦෼తͳΞΠσΟΞ͸;ͭ͏ʹ׆༻Ͱ͖ͦ͏
    • Influence ͕ϢʔβਪનʹͲͷ͘Β͍ޮ͔͘͸ϓϥοτϑΥʔϜͷੑ࣭ʹ
    ґଘͦ͠͏
    • interaction ͷछྨʹΑΔڧऑ͕ߟྀ͞Ε͍ͯͳ͍ͷ͕ؾʹͳͬͨ
    ঺հ࿦จ (1): Social Network Influence Ranking via Embedding Network Interactions for User Recommendation
    

    View Slide

12. ©2020 Wantedly, Inc.
    On Twitter Purge: A Retrospective Analysis of
    Suspended Users
    ঺հ࿦จຊ໨
    Farhan Asif Chowdhury , Lawrence Allen , Mohammad Yousuf , Abdullah Mueen
    

    View Slide

13. ©2020 Wantedly, Inc.
    Twitter ͕ abuse / span Ϣʔβͱͯ͠ purge ͨ͠Ϣʔβͨͪʹ
    য఺Λ౰ͯͯɼߦಈ΍ಛ௃Λ෼ੳ͍ͯ͠Δ
    ঺հ࿦จ (2): On Twitter Purge: A Retrospective Analysis of Suspended Users
    • author ͸ twitter ͷਓ͡Όͳ͍ͷͰσʔλ࡞Γ͔Βݥ͍͠…
    • Twitter API Λ༻͍ͯ͋Δ࣌఺ͰͷϢʔβϦετΛ࡞੒
    • 1 ϲ݄ޙʹ࠶౓ͦͷϢʔβͨͪΛऔಘ͢Δ͜ͱͰ
    purge ͞ΕͨϢʔβϦετΛ࡞͍ͬͯΔ
    • purge ͞Εͨޙ͔ΒͦͷϢʔβͷ activity Λऔಘ͢Δ͜ͱ͸Ͱ͖ͳ͍ͷͰ
    sampling Ͱͻͨ͢Β activity ΛूΊ͓͍ͯͯɼۮʑώοτͨ͠΋ͷΛ࢖͏
    

    View Slide

14. ©2020 Wantedly, Inc.
    key findings (ൈਮ)
    ঺հ࿦จ (2): On Twitter Purge: A Retrospective Analysis of Suspended Users
    • ࣗಈͰ࡞ΒΕΔ spam ͕ओ
    • ໨త͸੓࣏ܥ΍όΠϥϧϚʔέ͕ଟ͍
    • 60% Ҏ্͕ 2 ೥ؒ͸ Twitter Λ࢖͍ͬͯΔ (!)
    • ϓϩϑΟʔϧ͕ৗʹಉظ͍ͯ͠ΔΫϥελ͕͋ͬͨΓ͢Δ
    • purged Ϣʔβಉ͕࢜ RT ͋͠͏͜ͱ΋ଟ͍
    

    View Slide

15. ©2020 Wantedly, Inc.
    Friend or Faux:
    Graph-Based Early Detection of Fake Accounts
    on Social Networks
    ঺հ࿦จຊ໨
    Adam Breuer, Roee Eilat, Udi Weinsberg
    

    View Slide

16. ©2020 Wantedly, Inc.
    ໨తFake account Λૣظʹݕग़͍ͨ͠
    ঺հ࿦จ (3): Friend or Faux: Graph-Based Early Detection of Fake Accounts on Social Networks
    • Fake account ΛऔΓআ͘͜ͱ͸ɼSocial Networks ͷ݈શੑҡ࣋ʹ
    ͔ܽͤͳ͍ॏཁͳλεΫ
    • ͦͷѱӨڹ͕޿·Δ͜ͱΛ๷͙ͨΊʹ΋ɼૣظʹݕग़͢Δ͜ͱ͕ॏཁ
    • طଘͷΞϓϩʔν͸ߏங͞ΕͨωοτϫʔΫߏ଄ʹ஫໨͢Δ΋ͷ͕ଟ͘ɼ
    ωοτϫʔΫ͕े෼ʹߏங͞ΕΔ·Ͱݕग़͢Δ͜ͱ͕೉͔ͬͨ͠
    

    View Slide

17. ©2020 Wantedly, Inc.
    എܠͱલఏ
    • Facebook ͷݚڀ
    • 2019 1Q Ͱ 2.2 billion ΋ͷ fake account Λݕग़͍ͯ͠Δ…
    • طʹ 99.8% ͸Ϣʔβʹใࠂ͞ΕΔલʹݕग़Ͱ͖͍ͯΔ
    ঺հ࿦จ (3): Friend or Faux: Graph-Based Early Detection of Fake Accounts on Social Networks
    

    View Slide

18. ©2020 Wantedly, Inc.
    ՝୊: fake ϢʔβͷૹΔϦΫΤετ΋ͦͦ͜͜ Accept ͞Εͯ͠·͏
    ୯७ͳ Reject ཰ͷΑ͏ͳࢦඪͰ͸ݕ஌Ͱ͖ͳ͍
    ঺հ࿦จ (3): Friend or Faux: Graph-Based Early Detection of Fake Accounts on Social Networks
    

    View Slide

19. ©2020 Wantedly, Inc.
    Ұ෦ͷϢʔβ͸ɼfake Ϣʔβ͔ΒϦΫΤετΛૹΒΕ΍͍͢
    ʮfake Ϣʔβ͔ΒϦΫΤετΛૹΒΕ΍͍͢ϢʔβʯʹϦΫΤετΛૹ͍ͬͯΔ
    =
    fake Β͕͠͞ߴ͍ͱ͍͑ΔͷͰ͸ʁ
    ঺հ࿦จ (3): Friend or Faux: Graph-Based Early Detection of Fake Accounts on Social Networks
    

    View Slide

20. ©2020 Wantedly, Inc.
    Ұ෦ͷϢʔβ͸ɼreal or fake Λ൑ఆͯ͠ accept ͢Δ͔Λ൑அ͍ͯ͠Δ
    ʮfake Λڋ൱ͯ͠ real Λঝೝ͠΍͍͢ʯϢʔβ͔Βঝೝ͞Ε͍ͯΔ
    =
    real Β͕͠͞ߴ͍ͱ͍͑ΔͷͰ͸ʁ
    ঺հ࿦จ (3): Friend or Faux: Graph-Based Early Detection of Fake Accounts on Social Networks
    

    View Slide

21. ©2020 Wantedly, Inc.
    ఏҊख๏ͷ֓ཁ
    • ৽نϢʔβ͕୭ʹϦΫΤετΛૹ͍ͬͯΔ͔Λूܭ͢Δ
    • fake Ϣʔβ͔ΒϦΫΤετΛૹΒΕ΍͍͢Ϣʔβʹ஫໨͢Δ
    • ৽نϢʔβ͕ૹͬͨϦΫΤετ΁ͷฦ౴Λूܭ͢Δ
    • fake Ϣʔβ͔ΒͷϦΫΤετʹɼ;ͭ͏ͷϦΫΤετͱҧ͏൓ԠΛࣔ͢Α͏ͳϢʔβʹ஫໨͢Δ
    • ߴ͍֬৴౓Ͱ real or fake Λ༧ଌͰ͖ΔϢʔβʹ஫໨͢Δ
    ݸʑͷϢʔβ͝ͱͷಛ௃ΛՃຯͯ͠൑அ͢Δ͜ͱͰૣظݕ஌Λ࣮ݱ͢Δ
    ঺հ࿦จ (3): Friend or Faux: Graph-Based Early Detection of Fake Accounts on Social Networks
    

    View Slide

22. ©2020 Wantedly, Inc.
    ↑ʹՃ͑ͯ Confidence ͷߴ͍৘ใʹɼ
    ߴ͍ॏΈΛ༩͑ΔΑ͏ͳ re-weight Λߦ͏
    ʢ࣌ؒͷ౎߹Ͱলུ…ʣ
    SybilEdge
    ঺հ࿦จ (3): Friend or Faux: Graph-Based Early Detection of Fake Accounts on Social Networks
    

    View Slide

23. ©2020 Wantedly, Inc.
    ධՁ
    • ϦΫΤετ 5 ݅ͷஈ֊͔Βߴ͍ AUC
    • ୯७ͳ RejectRate Ͱ͸े෼ʹ৘ใ͕͋ͬͯ΋
    ݫ͍͜͠ͱ͕Θ͔Δ
    • false positive ΛݟͯΈΔͱɼreal ϢʔβͰ͸
    ͋Δ͕ਪનΛར༻ͯ͠େྔʹϦΫΤετΛૹΔ
    “spammy” ͳϢʔβͩͬͨ
    ঺հ࿦จ (3): Friend or Faux: Graph-Based Early Detection of Fake Accounts on Social Networks
    

    View Slide

24. ©2020 Wantedly, Inc.
    ·ͱΊ
    • fake ͔ΒϦΫΤετ͞Ε΍͍͢Ϣʔβ΍ɼ
    fake ͱ real Ͱঝೝ཰͕ҧ͏Ϣʔβ͕͍Δ͜ͱΛൃݟͨ͠
    • Ϣʔβ͝ͱͷ܏޲ʹ஫໨͢Δ͜ͱͰɼfake account ͷૣظݕग़Λվળͨ͠
    ײ૝
    • ࠷ऴతͳࣜ͸ώϡʔϦεςΟοΫͷੵΈ্͛Ͱɼ͜ΕͰ͜͜·Ͱͷੑೳ͕
    ग़ͤΔͷ͸͍͢͝
    • Fake account ଆ͕͔ͳΓຊؾ͡Όͳ͍ͱ͜͜·Ͱͷ෼ੳ͸ෆཁͳؾ͕͢Δ
    ͷͰɼFacebook ͱ͍͏ϓϥοτϑΥʔϜͷڧ͞Λײͨ͡
    ঺հ࿦จ (3): Friend or Faux: Graph-Based Early Detection of Fake Accounts on Social Networks
    

    View Slide