DEIM2023 大規模Webデータにおけるキーワード・地域ごとの拡散パターン抽出

Transcript

1. େن໛Webσʔλʹ͓͚Δ
   Ωʔϫʔυɾ஍Ҭ͝ͱͷ֦ࢄύλʔϯநग़
   ଜࢁଠҰ, দݪ༃ࢠ, ᓎҪอࢤ
   େࡕେֶ, ࢈ۀՊֶݚڀॴ
   DEIM 2023
   

   View Slide

2. 2
   ιʔγϟϧηϯαʔͷ໾ׂ
   എܠ
   ιʔγϟϧηϯαʔ ࣮ࡍͷࣾձ
   ౤ߘݕࡧ
   ϑΟʔυόοΫ
   

   View Slide

3. 3
   Google Flu
   എܠ
   

   View Slide

4. 4
   ݕࡧΫΤϦ͔Βજࡏతͳؔ܎Λଊ͑ΒΕΔ͔ʁ
   Ϟνϕʔγϣϯ
   U.S.
   Japan
   Italy
   Facebook
   Twitter
   Instagram
   Facebook
   Twitter Instagram
   Q: Ωʔϫʔυؒͷؔ܎͸ʁ
   Q: 国同⼠の関係は？
   ？
   

   View Slide

5. 5
   ໰୊ఆٛ
   (JWFO ΦϯϥΠϯ׆ಈσʔλ
   e.g., άʔάϧݕࡧ
   

   View Slide

6. 6
   ໰୊ఆٛ
   (PBM ஍Ҭؒͷ֦ࢄύλʔϯ
   e.g., άʔάϧݕࡧ
   taylor swift
   lady gaga
   beyonce katy perry
   maroon 5
   stevie wonder
   Group A: Russia, Mexico,
   Spain, Vietnam, etc.
   Group B: United Kingdom,
   Germany, France, etc.
   Group C: United States,
   Canada, Australia, etc.
   Group D: Iran, Pakistan,
   Brazil, Argentina, etc.
   Group E: China, Japan,
   Vietnam, South Africa, etc.
   

   View Slide

7. 7
   ໰୊ఆٛ
   (PBM Ωʔϫʔυؒͷؔ܎
   e.g., άʔάϧݕࡧ
   United States
   lady gaga maroon 5
   stevie
   wonder
   beyonce
   katy
   perry
   taylor
   swift
   Japan
   lady gaga maroon 5
   stevie
   wonder
   beyonce
   katy
   perry
   taylor
   swift
   

   View Slide

8. 8
   ໰୊ఆٛ
   (PBM قઅੑ
   e.g., άʔάϧݕࡧ
   

   View Slide

9. 9
   ໰୊ఆٛ
   (PBM ߴ͍༧ଌਫ਼౓
   e.g., άʔάϧݕࡧ
   2012 2014 2016 2018
   Time
   Keywords Countries
   Russia
   Mexico
   Colombia
   Modeling Forecasting
   beyonce
   katy perry
   stevie wonder
   taylor swift
   lady gaga
   maroon 5
   

   View Slide

10. 10
    ໰୊ఆٛ
    GivenΦϯϥΠϯ׆ಈσʔλ
    Goal 1 ஍Ҭؒͷ֦ࢄύλʔϯ
    Goal 2 Ωʔϫʔυؒͷؔ܎
    Goal 3 قઅੑ
    Goal 4ߴ͍༧ଌਫ਼౓
    

    View Slide

11. 11
    ໰୊ఆٛ
    Our solution: FLUXCUBE
    拡散反応⽅程式
    +
    ニューラルネットワーク
    GivenΦϯϥΠϯ׆ಈσʔλ
    Goal 1 ஍Ҭؒͷ֦ࢄύλʔϯ
    Goal 2 Ωʔϫʔυؒͷؔ܎
    Goal 3 قઅੑ
    Goal 4ߴ͍༧ଌਫ਼౓
    

    View Slide

12. 12
    ൓Ԡ֦ࢄํఔࣜ
    Our model: FLUXCUBE
    𝜕𝒖
    𝜕𝑡
    = 𝒇 𝒖, 𝑡 + 𝒟Δ𝑢
    Main Equation
    ൓Ԡ߲ ෼ࢄ߲
    Reaction-Diffusion System is utilized is
    a mathematical model corresponding
    for physical phenomena.
    

    View Slide

13. 13
    Our model: FLUXCUBE
    𝜕𝒖
    𝜕𝑡
    = 𝒇 𝒖, 𝑡 + 𝒟Δ𝑢
    Main Equation
    ෼ࢄ߲
    Amazon
    BestBuy
    ebay
    Target
    ڝ߹ؔ܎
    ૬ޓؔ܎
    Macy’s
    Washington
    State
    Amazon
    Target
    ebay
    Craigslist
    قઅ߲
    ϒϥοΫϑϥΠσʔ
    +
    ×
    ൓Ԡ֦ࢄํఔࣜ
    ൓Ԡ߲
    

    View Slide

14. 14
    Our model: FLUXCUBE
    ೖྗ: 𝝌 = 𝒙𝒕𝒊𝒋
    λΠϜεςοϓ U
    ஍Ҭ J ΞΠςϜ K
    𝜕𝒖
    𝜕𝑡
    = 𝒇 𝒖, 𝑡 + 𝒟Δ𝑢
    反応拡散⽅程式
    FLUXCUBE
    𝜕𝑥$%&
    𝜕𝑡
    = 𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    + ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕
    𝑥$)*%&
    = 𝐹
    𝜕𝑥$%&
    𝜕𝑡
    + 𝑥$%&
    ൓Ԡ߲ ֦ࢄ߲
    قઅ߲
    

    View Slide

15. 15
    Our model: FLUXCUBE
    𝜕𝒖
    𝜕𝑡
    = 𝒇 𝒖, 𝑡 + 𝒟Δ𝑢
    反応拡散⽅程式
    FLUXCUBE
    𝜕𝑥$%&
    𝜕𝑡
    = 𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    + ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕
    𝑥$)*%&
    = 𝐹
    𝜕𝑥$%&
    𝜕𝑡
    + 𝑥$%&
    ൓Ԡ߲
    Amazon
    BestBuy
    ebay
    Target
    Competitive relationship
    Mutualistic relationship
    Macy’s
    

    View Slide

16. 16
    ൓Ԡ߲: ΞΠσΞ
    ΞΠςϜͷؔ܎͸δϟϯάϧͷಈ෺ͷؔ܎ͱྨࣅ (Matsubara, 2015)
    ൓Ԡ߲: FLUXCUBE
    Spider
    monkeys Capybaras
    Squirrel
    monkeys
    Macaws
    Fruits Nuts Grass
    Facebook Twitter Youtube Reddit
    Kids
    Teens
    Adults
    類似
    

    View Slide

17. 17
    ൓Ԡ߲: ΞΠσΞ
    ϩτΧɾϘϧςϥํఔࣜ
    ൓Ԡ߲: FLUXCUBE
    Facebook Twitter Youtube Reddit
    Kids
    Teens
    Adults
    𝑎&
    𝝌𝒕,𝒊,𝒋
    1 −
    ∑&+
    𝑐&&+
    𝑥$%&+
    𝑏&
    

    View Slide

18. 18
    ൓Ԡ߲: ΞΠσΞ
    ϩτΧɾϘϧςϥํఔࣜ
    ൓Ԡ߲: FLUXCUBE
    Facebook Twitter Youtube Reddit
    Kids
    Teens
    Adults
    𝑎&
    𝝌𝒕,𝒊,𝒋
    1 −
    ∑&+
    𝑐&&+
    𝑥$%&+
    𝑏&
    𝒄𝒋𝒋
    : ΩʔϫʔυKͱΩʔϫʔυK`
    ؒͷ૬ޓ࡞༻ͷڧ͞
    

    View Slide

19. 19
    ൓Ԡ߲: ΞΠσΞ
    ϩτΧɾϘϧςϥํఔࣜ
    Reaction term: FLUXCUBE
    Facebook Twitter Youtube Reddit
    Kids
    Teens
    Adults
    𝑎&
    𝝌𝒕,𝒊,𝒋
    1 −
    ∑&+
    𝑐&&+
    𝑥$%&+
    𝑏&
    

    View Slide

20. 20
    ൓Ԡ߲: ΞΠσΞ
    ϩτΧɾϘϧςϥํఔࣜ
    ൓Ԡ߲: FLUXCUBE
    Facebook Twitter Youtube Reddit
    Kids
    Teens
    Adults
    𝑎&
    𝝌𝒕,𝒊,𝒋
    1 −
    ∑&+
    𝑐&&+
    𝑥$%&+
    𝑏&
    

    View Slide

21. 21
    ൓Ԡ߲: ΞΠσΞ
    ϩτΧɾϘϧςϥํఔࣜ
    ൓Ԡ߲: FLUXCUBE
    Facebook Twitter Youtube Reddit
    Kids
    Teens
    Adults
    𝑎&
    𝝌𝒕,𝒊,𝒋
    1 −
    ∑&+
    𝑐&&+
    𝑥$%&+
    𝑏&
    

    View Slide

22. 22
    ൓Ԡ߲: ΞΠσΞ
    Lotka-Volterra Equation Λ൓Ԡ߲ͱͯ͠ར༻
    ൓Ԡ߲: FLUXCUBE
    𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    = 𝑎&
    𝝌𝒕,𝒊,𝒋
    1 −
    ∑&+
    𝑐&&+
    𝑥$%&+
    𝑏&
    𝜕𝑥$%&
    𝜕𝑡
    = 𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    + ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕
    FLUXCUBE:
    

    View Slide

23. 23
    ֦ࢄ߲: FLUXCUBE
    𝜕𝒖
    𝜕𝑡
    = 𝒇 𝒖, 𝑡 + 𝒟Δ𝑢
    反応拡散⽅程式
    FLUXCUBE
    𝜕𝑥$%&
    𝜕𝑡
    = 𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    + ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕
    𝑥$)*%&
    = 𝐹
    𝜕𝑥$%&
    𝜕𝑡
    + 𝑥$%&
    Seasonal term
    ֦ࢄ߲
    Washington
    State
    Amazon
    Target
    ebay
    Craigslist
    

    View Slide

24. 24
    ֦ࢄ߲: ΞΠσΞ
    l Ωʔϫʔυؒͷ૬ޓ࡞༻͸֎෦ཁҼͳͲʹΑΓ
    ҰఆͰ͸ͳ͍
    l ෳࡶͳݱ৅΍ٸܹͳมԽͳͲʹΑͬͯੜ͡Δɼ
    ࣌ؒมԽ͢Δ૬ޓ࡞༻Λଊ͑Δඞཁ
    ֦ࢄ߲: FLUXCUBE
    Washington
    State
    Amazon
    Target
    ebay
    Craigslist
    

    View Slide

25. 25
    ֦ࢄ߲: ΞΠσΞ
    l Ωʔϫʔυؒͷ૬ޓ࡞༻͸֎෦ཁҼͳͲʹΑΓ
    ҰఆͰ͸ͳ͍
    l ෳࡶͳݱ৅΍ٸܹͳมԽͳͲʹΑͬͯੜ͡Δɼ
    ࣌ؒมԽ͢Δ૬ޓ࡞༻Λଊ͑Δඞཁ
    ֦ࢄ߲: FLUXCUBE
    Washington
    State
    Amazon
    Target
    ebay
    Craigslist
    相互作⽤はPhysics-informed neural
    networksのアイデアを活⽤して
    RNNを利⽤
    

    View Slide

26. 26
    Physics-informed neural Network (PINN)
    l ෺ཧϞσϧͷؔ਺͸ਂ૚ֶशͰۙࣅՄೳ
    l ॱ໰୊: ํఔࣜͷۙࣅղ
    l ٯ໰୊: ؍ଌσʔλ͔ΒύϥϝʔλΛਪఆ
    l ༩͑ΒΕֶͨशσʔλͱɺٻΊ͍ͨύϥϝʔλ͕ํఔࣜΛຬ
    ͨ͢Α͏ʹղۭؒΛ੍ݶ͢Δ͜ͱͰద੾ͳύϥϝʔλΛਪఆ
    ֦ࢄ߲: FLUXCUBE
    ภඍ෼ํఔࣜͰ͋Δ֦ࢄ൓Ԡํఔࣜͷ
    ύϥϝʔλͷҰ෦Λ/FVSBM/FUXPSL
    ͰॊೈʹදݱΛࢼΈΔ
    

    View Slide

27. 27
    ֦ࢄ߲: FLUXCUBE
    ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕 = 8 𝑅𝑁𝑁(1: 𝑡) ⊙ 𝝌𝒕
    𝜕𝑥$%&
    𝜕𝑡
    = 𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    + ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕
    FLUXCUBE:
    l ૬ޓ࡞༻ͷਪఆʹRecurrent Neural NetworkΛར༻
    l ॱ໰୊: ํఔࣜͷۙࣅղ
    l ٯ໰୊: ؍ଌσʔλ͔ΒύϥϝʔλΛਪఆ
    l ਺ཧϞσϧͷҰ෦ʹχϡʔϥϧωοτϫʔΫΛ׆༻͢Δ͜ͱ
    ͰॊೈͳϞσϦϯάͱߴ͍આ໌ੑΛཱ྆
    

    View Slide

28. 28
    RNNͷग़ྗ஋ʹΑͬͯɺ֤Ωʔϫʔυͷ֦ࢄͷӨڹΛදݱ
    ֦ࢄ߲: FLUXCUBE
    ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕 = 8 𝑅𝑁𝑁(1: 𝑡) ⊙ 𝝌𝒕
    𝑹𝑵𝑵(𝟏: 𝒕)
    Time
    カナダからフランスへ
    の各キーワードの影響
    の流れ
    

    View Slide

29. 29
    قઅ߲: FLUXCUBE
    𝜕𝒖
    𝜕𝑡
    = 𝒇 𝒖, 𝑡 + 𝒟Δ𝑢
    拡散反応⽅程式
    FLUXCUBE
    𝜕𝑥$%&
    𝜕𝑡
    = 𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    + ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕
    𝑥$)*%&
    = 𝐹
    𝜕𝑥$%&
    𝜕𝑡
    + 𝑥$%&
    قઅ߲
    Washington
    State
    Amazon
    Target
    ebay
    Craigslist
    

    View Slide

30. 30
    ΦϯϥΠϯσʔλ͸ΫϦεϚε΍ϒϥοΫϑϥΠσʔͳͲ
    ͷقઅΠϕϯτʹԠͯ͡มԽ
    قઅ੒෼ͷநग़
    قઅ߲: FLUXCUBE
    𝑥$)*%&
    = 𝐹 ,-!"#
    ,$
    + 𝑥$%&
    = 1 + 𝑺$ ./0 1 %&
    ⊙ ,-!"#
    ,$
    + 𝑥$%&
    季節性の周期
    e.g., p = 52週間
    ϒϥοΫϑϥΠσʔ
    

    View Slide

31. 31
    Our model: FLUXCUBE
    𝜕𝒖
    𝜕𝑡
    = 𝒇 𝒖, 𝑡 + 𝒟Δ𝑢
    反応拡散⽅程式
    FLUXCUBE
    𝜕𝑥$%&
    𝜕𝑡
    = 𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    + ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕
    𝑥$)*%&
    = 𝐹
    𝜕𝑥$%&
    𝜕𝑡
    + 𝑥$%&
    ೖྗ: 𝝌 = 𝒙𝒕𝒊𝒋
    λΠϜεςοϓ U
    ஍Ҭ J ΞΠςϜ K
    ൓Ԡ߲ ֦ࢄ߲
    قઅ߲
    

    View Slide

32. 32
    Fitting: FLUXCUBE
    Loss Function 𝝌𝒄 − @
    𝝌 𝟐 + 𝛼 8 𝐷 4 + 𝛽 8 𝑆 4
    ճؼ߲ εύʔε߲
    𝜕𝑥$%&
    𝜕𝑡
    = 𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    + ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕
    𝑥$)*%&
    = 𝐹
    𝜕𝑥$%&
    𝜕𝑡
    + 𝑥$%&
    λΠϜεςοϓ U
    ஍Ҭ J ΞΠςϜ K
    ೖྗ: 𝝌 = 𝒙𝒕𝒊𝒋
    

    View Slide

33. 33
    l ଟ͘ͷ஍Ҭؒͷ૬ޓ࡞༻Λਪ࿦͢Δ͜ͱ͸ɺܭࢉίετ͕ߴ͍
    ஍ҬͷάϧʔϓΠϯά
    

    View Slide

34. 34
    l ଟ͘ͷ஍Ҭؒͷ૬ޓ࡞༻Λਪ࿦͢Δ͜ͱ͸ɺܭࢉίετ͕ߴ͍
    ஍ҬͷάϧʔϓΫϥελϦϯά
    解決策:
    学習前に似た地域を
    １つのグループとして扱う
    ΞϝϦΧ ೔ຊ
    ࣅͨ
    ૬ޓ࡞༻
    ಉ͡άϧʔϓ
    

    View Slide

35. 35
    ஍ҬͷάϧʔϓΫϥελϦϯά
    𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    = 𝑎&
    𝝌𝒕,𝒊,𝒋
    1 −
    ∑&+
    𝑐&&+
    𝑥$%&+
    𝑏&
    𝜕𝑥$%&
    𝜕𝑡
    = 𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    + ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕
    FLUXCUBE:
    𝑔*, … 𝑔5$
    =K-means UMAP 𝒂𝒊, 𝒃𝒊, 𝑪𝒊 , 𝑖 = [1, … , 𝐿]
    各グループ
    を推定
    各地域の反応項のパラメータ
    地域の数
    ൓Ԡ߲
    

    View Slide

36. 36
    ஍ҬͷάϧʔϓΫϥελϦϯά
    𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    = 𝑎&
    𝝌𝒕,𝒊,𝒋
    1 −
    ∑&+
    𝑐&&+
    𝑥$%&+
    𝑏&
    𝜕𝑥$%&
    𝜕𝑡
    = 𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    + ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕
    FLUXCUBE:
    𝑔*, … 𝑔5$
    =K-means UMAP 𝒂𝒊, 𝒃𝒊, 𝑪𝒊 , 𝑖 = [1, … , 𝐿]
    Reaction term
    taylor swift
    lady gaga
    beyonce katy perry
    maroon 5
    stevie wonder
    Group A: Russia, Mexico,
    Spain, Vietnam, etc.
    Group B: United Kingdom,
    Germany, France, etc.
    Group C: United States,
    Canada, Australia, etc.
    Group D: Iran, Pakistan,
    Brazil, Argentina, etc.
    Group E: China, Japan,
    Vietnam, South Africa, etc.
    最適なグループ数 (𝑑!)を
    Model Description Cost
    (MDL) を⽤いて探索
    各グループ
    を推定
    各地域の反応項のパラメータ
    地域の数
    

    View Slide

37. 37
    ࣮ݧઃఆ
    l 13, 26, 52 िؒઌͷ༧ଌ஋ͷਫ਼౓ΛධՁ
    l ධՁࢦඪ
    l RMSE, MAE: খ͍͞஋Ͱ͋Δ΄Ͳߴ͍ਫ਼౓
    σʔληοτ
    l Google Trend͔ΒऔಘՄೳͳ10೥ؒ
    ͷ1िؒ͝ͱͷݕࡧ਺ (7छྨ)
    ࣮ݧ
    

    View Slide

38. 38
    ࣮ݧ݁Ռ
    ࣮ݧ
    

    View Slide

39. 39
    ࣮ݧ݁Ռ
    ࣮ݧ
    

    View Slide

40. 40
    ࣮ݧ݁Ռ
    ࣮ݧ
    

    View Slide

41. 41
    Ablation Study
    ࣮ݧ
    𝜕𝑥$%&
    𝜕𝑡
    = 𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    + ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕
    𝑥$)*%&
    = 𝐹
    𝜕𝑥$%&
    𝜕𝑡
    + 𝑥$%&
    قઅ߲Λ࡟আ
    ֦ࢄ߲Λ࡟আ
    ֦ࢄ߲Λఆ਺ʹஔ͖׵͑
    ఏҊϞσϧ
    

    View Slide

42. 42
    Ablation Study
    ࣮ݧ
    𝜕𝑥$%&
    𝜕𝑡
    = 𝒇 𝑥$%&
    |𝝌𝒕,𝒊,:
    + ℊ 𝑥$%&
    |𝝌𝒕
    , 𝒕
    𝑥$)*%&
    = 𝐹
    𝜕𝑥$%&
    𝜕𝑡
    + 𝑥$%&
    ℊ 𝑥)*+|𝝌𝒕 , 𝒕 = + 𝑅𝑁𝑁(1: 𝑡) ⊙ 𝝌𝒕
    𝜶 ⊙ 𝝌𝒕
    قઅ߲Λ࡟আ
    ֦ࢄ߲Λ࡟আ
    ֦ࢄ߲Λఆ਺ʹஔ͖׵͑
    ఏҊϞσϧ
    

    View Slide

43. 43
    Case study1: Vod
    Case Study
    AppleTV / ESPN / HBO / Hulu / Netflix
    Sling / Vudu / Youtube
    Keyword List
    

    View Slide

44. 44
    Case study1: Vod
    Case Study
    ૬ޓ࡞༻
    

    View Slide

45. 45
    Case study1: Vod
    Case Study
    影響の流れ
    

    View Slide

46. 46
    Case study2: SNS
    Case Study
    Line/facebook/slack/snapchat/
    twitter/viber/whatsapp
    Keyword List
    

    View Slide

47. 47
    Case study2: SNS
    Case Study
    ૬ޓ࡞༻
    

    View Slide

48. 48
    Case study2: SNS
    Case Study
    影響の流れ
    

    View Slide

49. 49
    l FLUXCUBE: ൓Ԡ֦ࢄํఔࣜͱੜଶܥΛදݱ͢Δ਺ཧϞσϧ -PULB
    7PMUFSB NPEFM
    ʹجͮ͘ϞσϦϯάख๏ɽೖྗ͞Εͨσʔλʹજࡏత
    ʹଘࡏ͢Δقઅੑ΍૬ޓ࡞༻ͳͲͷྗֶతͳؔ܎Λநग़
    l (PPHMF5SFOEΛ༻͍ͨσʔλΛ༻͍ͯɼఏҊϞσϧͷ༗ޮੑΛఏࣔ
    l ؍ଌσʔλͷഎޙʹଘࡏ͢Δજࡏతͳ૬ޓ࡞༻΍ӨڹͷྲྀΕΛɼਓؒ
    ͕ղऍՄೳͳܗΛఏڙ
    ࠓޙͷల๬
    l ࣗಈతʹؔ܎ͷ͋ΔΩʔϫʔυͱؔ܎ͷͳ͍ΩʔϫʔυΛ۠ผ͢Δ͜
    ͱΛݕ౼
    ·ͱΊ
    

    View Slide