Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Kaggle日記について

fkubota
June 02, 2021

 Kaggle日記について

Kaggle日記をなぜ作ったのかとか話してます。

Kaggle日記の詳細はこちら↓

Kaggle日記という戦い方: https://zenn.dev/fkubota/articles/3d8afb0e919b555ef068

fkubota

June 02, 2021
Tweet

More Decks by fkubota

Other Decks in Technology

Transcript

  1. Kaggle೔هʹ͍ͭͯ
    ྺ࢙ͱ͔എܠͱ͔
    fkubota

    View Slide

  2. ࣗݾ঺հ 02
    fkubota (Twitter, Kaggle)
    - όϯυϧΧʔυͷձࣾ(ΧϯϜ)ͰػցֶशΤϯδχΞ
    - Kaggle Expert
    - ԭೄग़਎(౦ژʹग़͖ͯͯ3೥൒)
    - ෺ཧֶՊग़਎(ڧ૬ؔిࢠܥɺ4.2KͷӷମϔϦ΢Ϝʹ࣓ੑମಥͬࠐΜͰͨ)
    - ϓϩάϥϛϯάྺ͸2೥൒͙Β͍
    - झຯ
    - ૣى͖(4࣌൒ىচ)
    - Kaggle
    - ίʔώʔɺϏʔϧɺ΢ΟεΩʔ
    - ಡॻɺ෺ཧɺ఩ֶ

    View Slide

  3. ࠓ೔࿩͢͜ͱ
    Kaggle೔هͬͯԿʁ ॳظʹ
    ίϯϖਏ͔ͬͨ࿩
    ͔ͩΒ
    Kaggle೔ه࢝Ίͨ
    ͱΓ͋͑ͣͲΜͳ΋ͷ͔
    ؆୯ʹ঺հ͠·͢ɻ
    هࣄ͕͋ΔΑɻ(url)
    ίϯϖ࢝Ίͨࠒɺશવ΍
    Δؾى͖ͳ͔ͬͨͷͰɺ
    ͦͷ๊͍࣌͑ͯͨ໰୊Λ
    ͪΐͬͱ੔ཧɻ
    Kaggle೔هΛ࢝Ίͨཧ༝Λ
    ఻͑Ε͹ΑΓ࢖͍΍͘͢ͳ
    Δ͔ͳͱɻ
    ͦ΋ͦ΋Kaggle೔هͰ͋Δ
    ඞཁ͢Βͳ͍ͱ͍͏࿩ɻ
    ओʹϏΪφʔʹ޲͚ͯʂ
    03

    View Slide

  4. Kaggle೔هͬͯԿʁ

    View Slide

  5. ͜Μͳײ͡ʂ
    Kaggle೔ه
    ࣮ݧܭը
    ࣮ݧ݁Ռͱߟ࡯
    ࿦จ/ࢀߟจݙ
    Discussion
    KaggleCode
    ΞΠσΞ
    KaggleࢀՃதʹੜ·ΕΔ৘ใ͸શͯKaggle೔هʹ
    ྫ) ௗίϯϖͷKaggle೔ه
    05

    View Slide

  6. Kaggleͷେมͳͱ͜
    difficult
    toooooo long
    Կ͔ͯͨ͠๨ΕͪΌͬͨ
    ໋໊نଇͷഁ୼
    Ϟνϕଓ͔ͳ͍Α...
    ΞΠσΞͷރׇ
    too long໰୊΁ͷରࡦ
    06

    View Slide

  7. Kaggle೔ه͕औΓ૊ΜͰ͍Δ͜ͱ
    Ϟνϕҡ࣋ ϑΝΠϧ໊Ͱ؅ཧ͠ͳ͍
    ೴ͷϝϞϦͰ؅ཧ͠ͳ͍ τϨʔαϏϦςΟ
    ৄࡉ͸ʮKaggle೔هͱ͍͏ઓ͍ํʯΛݟ͍ͯͩ͘͞m(_ _)m
    ௕͍ͷʹ΄ͱΜͲ͏·͍͔͘ͳ
    ͍ͱ͔஍ࠈ͡ΌͶʔ͔ɻ
    Ͳ͏ͤഁ୼͢ΔΜͰ͠ΐʁ
    ͳΒ࢝Ί͔Β؅ཧ͢Δ͜ͱΛ͋
    ͖ΒΊΑ͏ͥɻ
    3ϲ݄΋͋ΔΜͩΑʁ
    ֮͑ΒΕΔΘ͚͕ͳ͍ɻ
    ݴ͍͍͚ͨͩɻ
    07

    View Slide

  8. ॳظʹίϯϖਏ͔ͬͨ࿩

    View Slide

  9. Kaggleʹڵຯ࣋ͪ࢝Ίͨ࣌
    ڵຯ
    ਓ޻஌ೳΛڝ͍͍͋
    ͳ͕Βֶ΂Δͩͱʂʁ
    ΍Δ͔͠Ͷ͐ʂ
    ׬શཧղ
    ܾఆ໦ͱSVMΛཧղͨ͠ɻ
    ͑ʁsklearnͰ͙͢࢖͑Δͷʁ
    ༨༟͡ΌΜww
    ੩ऐ
    ܾఆ໦ͱSVMࢼ͔ͨ͠
    Β΋͏΍Δ͜ͱͳ͍Αʁ
    1αϒͰऴྃʂ
    σΟεΧογϣϯ΋Կॻ͍ͯΔ
    ͔Θ͔ΒΜ͜͠ΕҎ্͸ਏ͍ʂ
    09

    View Slide

  10. ޛͬͨ๻
    ·ͩ...ͦͷ࣌Ͱ͸ͳ͍
    10

    View Slide

  11. ۭനͷ2ϲ݄
    ͦͷ͕࣌๚ΕΔͷΛษڧ͠ͳ͕Β଴ͪ·ͨ͠
    11

    View Slide

  12. ຊ౰ʹޛͬͨ๻
    ͦͷ࣌ ͸Ұੜ΍ͬͯ͜ͳ͍
    12

    View Slide

  13. ࣮ફ͸΍ͬͺҧ͏ΜͩΖ͏ͳ͊
    ͓ษڧ ࣮ફ
    ࣮ફͰֶ͔͠΂ͳ͍͜ͱͬͯ͋ΔΑͶ͐
    - EDA
    - ϦʔΫ
    - CVͷ੾Γํ
    - ಛ௃ྔબ୒
    - ͳͲͳͲ...
    Θ͔Δ͚Ͳ͠ΜͲ͍΋Μ͸͠ΜͲ͍
    ޲͍ͯͳ͍ͷ͔ͳʁ
    13

    View Slide

  14. ίϛοτྔ
    ָ͠͞
    Կֶ͔Ϳ͍͍࣌ͬͯͩͨ͜Μͳײ͡ۂઢ
    0
    φχίϨ?
    φχίϨ?
    φχίϨ?
    ݴޠԽΉ͔͍͚ͣ͠Ͳ...
    - ஌ࣝͷମܥԽ͕࢝·Δ?
    - ఺ͱ఺͕ܨ͕Γ࢝ΊΔ?
    - Θ͔Βͳ͍͕Θ͔Δ?
    - ήʔϜͷϧʔϧ͕Θ͔Δʁ
    - ….Έ͍ͨͳײ͡ʁ
    ͋͘·Ͱ๻ͷܦݧྫͳΜͰ͕͢ 14

    View Slide

  15. Կ͕ݴ͍͍͔ͨͱݴ͏ͱ
    ָ͠͞
    0
    ͻͱ·ͣ͜͜ΛͲ͏ʹ͔
    ৐Γӽ͑Α͏ͱࢥͬͨ
    ͜͜Ͱ޲͍͍ͯͳ͍ͱ൑அ͢Δͷ
    ͸΋͍ͬͨͳ͍͔ͳ͊
    ͋͘·Ͱ๻͕ͦ͏ࢥ͚ͬͨͬͯͩͶʂʂ
    15
    φχίϨظ

    View Slide

  16. ໰୊͕গ͠໌֬ʹͳͬͯ
    ίϯϖ͠ΜͲ͍
    Ͳ͏͠Α͏໰୊
    16
    φχίϨظΛ
    Ͳ͏৐Γӽ͑Δ͔໰୊

    View Slide

  17. ͔ͩΒKaggle೔ه͸͡Ίͨ

    View Slide

  18. ετϨεʹͳΓͦ͏ͳ΋ͷ
    ໋໊نଇͷഁ୼ ࡞ۀ࠶։ίετ ΞΠσΞͷރׇ
    ࢀߟจݙͷ؅ཧ
    outputϑΝΠϧͷ؅ཧ
    ࠶ݱੑ͕ͳ͍
    18
    2೔ۭ͚ͪΌͬͯ࠶։͠Α
    ͏ͱࢥ͚ͬͨͲΊΜͲ͘͞
    ͍͔Β໌೔΍Ζ͏ɻ
    ΋͏ΞΠσΞ͕ਚ͖·ͨ͠ɻ
    A.pyಈ͔ͯ͠Ͱ͖ͨa.csv
    Λ࢖ͬͯB.pyΛಈ͔ͤ͹࠶
    ݱͰ͖Δ͸ͣͩΑɻ
    ·͋ɺͰ͖ͳ͔ͬͨΜͰ͢
    ͚ͲͶ:)
    ࢀߟʹͳΓͦ͏ͳ࿦จ΍Βه
    ࣄ΍ΒΛอଘ͚ͨ͠ͲͲΕΛ
    ಡΜͰͲΕΛಡΜͰͳ͍ͷ͔
    Θ͔ΒΜɻ
    exp_5fold.ipynb
    exp_6fold_3.ipynb
    exp_5fold_seed42.ipynb
    exp_5fold_nn_v5.ipynb
    exp_4fold_nn_prepro.ipynb
    exp_5fold_nn_postproc_v2.ipynb
    exp_5fold_fix_bug.ipynb
    ແݶϧʔϓ
    ͜ͷ
    ultra_super_feature.csv
    ͸Ͳ͏΍ͬͯ࡞ΒΕͨΜͰ͢
    ͔Ͷʁ
    Todo Doing Done

    View Slide

  19. ରࡦͰ͖Ε͹ԿͰ΋ྑ͍
    લϖʔδʹ্͛ͨΑ͏ͳ໰୊ʹͰ͖Δ͚ͩແཧͳ͘ରॲ͢Δ
    ͨΊʹKaggle೔هͱ͍͏ํ๏ΛऔΓ·ͨ͠
    ͨͩɺͪΐͬͱ஫ҙͳͷͰ͕͢
    - Kaggle೔هͰ͋Δඞཁ͸ͳ͍͠
    - ͦ΋ͦ΋ࡉ͔͍໰୊ʹͦΕͧΕରॲΛແཧͯ͢͠Δඞཁ
    ΋ͳ͍͠
    - ͳΜͳΒ͠ΜͲ͍ͳΒKaggle΍Βͳ͍બ୒ࢶ΋͋Δ
    ͷͰ๻ͷҙݟ͸બ୒ࢶͷҰ͙ͭΒ͍ʹड͚औ͍ͬͯͩ͘͞
    19

    View Slide

  20. ෼ࢠίϯϖͷࢥ͍ग़ ॳࢀઓͷίϯϖʂʂ
    Kaggle೔هσϏϡʔʂʂ
    ࣌ؒ
    public LB
    (score)
    0 pandasͬͯ
    φϯμεʁ
    3ϲ݄
    ಔϝμϧϥΠϯ
    groupby
    ŧŔŕŪżŞƂŜŽūŘ
    oofͬͯԿʂʁ
    lightGBMʁʁʁ
    ςʔϒϧ͕ͳΜͰ
    ͨ͘͞Μ͋Δͷʂʁ
    ٙ໰͕ͨ·ͬͨࠒʹͨ·ͨ·
    ΧϨʔ͞Μͷ
    ʮKaggleͷνϡʔτϦΞϧʯ
    ʹग़ձ͍׬શʹཧղ͢Δ
    ࢒Γ2೔͙Β͍
    ͰॳΊͯ
    ಔϝμϧݍ಺ʹ
    ͜ͷ࣌ظ͸codeͱdiscussionΛશ෦ݟͯͱΓ͋
    ͑ͣ࢖͏ͱ͍͏ͷΛߦ͍είΞ্͕͕ͬͨɻ
    ͨͩެ։ϊʔτϒοΫະຬɻ
    ͜͜·ͰஷΊ͍ͯͨΦϦδφϧͷ
    ΞΠσΞͷ࣮૷Λ͸͡Ίͨɻ
    ͱʹ͔࣮͘૷଎౓͕஗͍ɻ
    15ݸఔ౓ࢼͯ͠
    3ݸ͙Β͍͕౰ͨͬͨɻ
    ࠷ޙͷ7೔ͰΞϯαϯϒ
    ϧΛษڧͯ͠ࢼͨ͠ɻ
    ಔϝμϧ(9%)
    ϑΟχογϡʂ
    20

    View Slide

  21. Thanks :)

    View Slide