機械学習と解釈可能性

@Tokyo.R #67
https://atnd.org/events/93829

Transcript

1. ػցֶशͱղऍՄೳੑ
   Masaaki Horikoshi @ ARISE analytics
   

   View Slide

2. ࣗݾ঺հ
   • R
   • ύοέʔδ։ൃͳͲ
   • Git Awards ࠃ಺1Ґ
   • Python
   • http://git-awards.com/users/search?login=sinhrks
   

   View Slide

3. Interpretability
   ղऍՄೳੑ
   

   View Slide

4. ղऍՄೳੑͱ͸
   • ໌֬ͳఆٛ͸ͳ͍͕
   • Ϟσϧͷग़ྗ (what) ͚ͩͰͳ͘ɺͦͷཧ༝ (why) Λઆ໌͢Δ͜ͱ
   • ͦͷͨΊʹɺͳΜΒ͔ͷख๏/ج४Λ༻͍ͯϞσϧΛղऍ͢Δ͜ͱ
   The Mythos of Model Interpretability (Lipton, 2016)
   

   View Slide

5. ͳʹ͕خ͍͔͠
   • Trust
   • Causality
   • Transferability
   • Informativeness
   • Fair and Ethical Decision Making
   The Mythos of Model Interpretability (Lipton, 2016)
   

   View Slide

6. ղऍͷͨΊͷΞϓϩʔν
   1. આ໌͠΍͍͢ػցֶशख๏ΛબͿ
   • ਫ਼౓͕ෆे෼ͳ৔߹͕͋Δ
   2. ͳΜΒ͔ͷղऍख๏Λ࢖͏
   

   View Slide

7. ղऍՄೳੑ
   • Global Interpretability
   • Ϟσϧ΍σʔλશମͷ܏޲Λղऍ
   • ۙࣅ΍ཁ໿౷ܭྔΛར༻ => ہॴతʹ͸ෆਖ਼֬ͳ৔߹΋
   • Local Interpretability
   • Ϟσϧ΍σʔλͷݶΒΕͨྖҬΛղऍ
   • ΑΓਖ਼֬ͳઆ໌͕Մೳ
   

   View Slide

8. ղऍՄೳੑ
   • ద੾ͳख๏͸ʮԿΛʯղऍ͍͔ͨ͠ʹґଘ
   .PEFM4QFDJpD .PEFM"HOPTUJD
   (MPCBM
   *OUFSQSFUBCJMJUZ
   w 3FHSFTTJPO$PF⒏DJFOUT
   w 'FBUVSF*NQPSUBODF
   ʜ
   w 4VSSPHBUF.PEFMT
   w 4FOTJUJWJUZ"OBMZTJT
   ʜ
   -PDBM
   *OUFSQSFUBCJMJUZ
   w .BYJNVN"DUJWBUJPO"OBMZTJT
   ʜ
   w -*.&
   w -0$0
   w 4)"1
   ʜ
   

   View Slide

9. Regression Coefficients
   • ඪ४Խภճؼ܎਺
   library(dplyr)
   library(mlbench)
   data(BostonHousing)
   df <- BostonHousing %>%
   mutate_if(is.factor, as.numeric) %>%
   mutate_at(-14, scale)
   head(df, 3)
   crim zn indus chas nox rm age
   1 -0.4193669 0.2845483 -1.2866362 -0.2723291 -0.1440749 0.4132629 -0.1198948
   2 -0.4169267 -0.4872402 -0.5927944 -0.2723291 -0.7395304 0.1940824 0.3668034
   3 -0.4169290 -0.4872402 -0.5927944 -0.2723291 -0.7395304 1.2814456 -0.2655490
   dis rad tax ptratio b lstat medv
   1 0.140075 -0.9818712 -0.6659492 -1.4575580 0.4406159 -1.0744990 24.0
   2 0.556609 -0.8670245 -0.9863534 -0.3027945 0.4406159 -0.4919525 21.6
   3 0.556609 -0.8670245 -0.9863534 -0.3027945 0.3960351 -1.2075324 34.7
   

   View Slide

10. Regression Coefficients
    • ඪ४Խภճؼ܎਺
    • coefplot ύοέʔδͰภճؼ܎਺ͷՄࢹԽ
    ͕Մೳ
    lm.fit <- lm(medv ~ ., data = df)
    coef(lm.fit)
    (Intercept) crim zn indus chas nox rm
    22.53280632 -0.92906457 1.08263896 0.14103943 0.68241438 -2.05875361 2.67687661
    age dis rad tax ptratio b lstat
    0.01948534 -3.10711605 2.66485220 -2.07883689 -2.06264585 0.85010886 -3.74733185
    

    View Slide

11. • coefplot
    Regression Coefficients
    library(coefplot)
    coefplot(lm.fit)
    σʔλͷෆඋΛ౷ܭతʹݟൈ͘ (Gelman's Secret Weapon)
    http://d.hatena.ne.jp/hoxo_m/20150617/p1
    

    View Slide

12. • Feature Importance (Random Forest)
    Feature Importance
    library(dplyr)
    library(mlbench)
    data(Sonar)
    head(Sonar, 3)
    V1 V2 V3 V4 V5 V6 V7 V8 V9
    1 0.0200 0.0371 0.0428 0.0207 0.0954 0.0986 0.1539 0.1601 0.3109
    2 0.0453 0.0523 0.0843 0.0689 0.1183 0.2583 0.2156 0.3481 0.3337
    3 0.0262 0.0582 0.1099 0.1083 0.0974 0.2280 0.2431 0.3771 0.5598
    …
    V55 V56 V57 V58 V59 V60 Class
    1 0.0072 0.0167 0.0180 0.0084 0.0090 0.0032 R
    2 0.0094 0.0191 0.0140 0.0049 0.0052 0.0044 R
    3 0.0180 0.0244 0.0316 0.0164 0.0095 0.0078 R
    

    View Slide

13. • Feature Importance (Random Forest)
    Feature Importance
    library(caret)
    ca.fit <- train(Class ~ ., data = Sonar, method = "rf", ntree = 100)
    varImp(ca.fit)
    rf variable importance
    only 20 most important variables shown (out of 60)
    Overall
    V11 100.00
    V48 96.25
    V45 75.45
    V13 71.74
    V10 60.50
    …
    

    View Slide

14. • Feature Importance (Random Forest)
    Feature Importance
    plot(varImp(ca.fit), top = 10)
    

    View Slide

15. • caretͷvarImp͸ɺطఆͰख๏ґଘͷܭࢉํ๏Λར༻
    • help(varImp)
    Feature Importance
    randomForest::importance(ca.fit$finalModel) %>%
    as.data.frame %>%
    tibble::rownames_to_column(var = 'variable') %>%
    arrange(desc(MeanDecreaseGini)) %>%
    head(20)
    varImp(ca.fit, scale = F)
    Overall
    V11 3.793
    V48 3.684
    V45 3.078
    V13 2.970
    V10 2.643
    …
    variable MeanDecreaseGini
    1 V11 3.792926
    2 V48 3.683850
    3 V45 3.077996
    4 V13 2.970039
    5 V10 2.642881
    …
    

    View Slide

16. Feature Importance
    • randomForest::importance Ͱ͸ҎԼͷ૊Έ߹
    ΘͤͰಛ௃ྔͷॏཁ౓Λܭࢉ
    $MBTTJpDBUJPO 3FHSFTTJPO
    .FBOEFDSFBTFJOBDDVSBDZDPNQVUFE
    GSPNQFSNVUJOH00#EBUB
    &SSPS3BUF
    .FBO%FDSFBTF"DDVSBDZ
    
    .4&
    *OD.4&
    
    .FBOEFDSFBTFJOOPEFJNQVSJUZ
    (JOJJOEFY
    .FBO%FDSFBTF(JOJ
    
    344
    *OD/PEF1VSJUZ
    
    

    View Slide

17. OOB
    αϯϓϦϯά
    (෮ݩநग़)
    …
    OOB
    ࢀߟ: OOBσʔλ͔Βͷಛ௃ྔ
    ͷॏཁ౓ͷܭࢉ
    ֶश
    ֶश
    OOB
    ༧ଌ
    OOB
    ༧ଌ
    …
    ֤ྻͷ஋Λγϟοϑϧ͠ɺOOBޡΓ཰ͱൺֱ
    ༧ଌ ൺֱ
    …
    ༧ଌ ൺֱ
    

    View Slide

18. • caret ͸Ϟσϧґଘ͠ͳ͍ॏཁ౓΋αϙʔτ
    Feature Importance
    (Model Agnostic)
    varImp(ca.fit, useModel = F)
    ROC curve variable importance
    only 20 most important variables shown (out of 60)
    Importance
    V11 100.00
    V12 86.30
    V10 82.64
    V49 82.12
    V9 81.97
    …
    

    View Slide

19. • Classification
    • આ໌ม਺͝ͱʹROC෼ੳ
    • ଟΫϥεͷ৔߹͸2Ϋϥεͷ૊Έ߹Θͤ͝ͱ
    • Regression
    • આ໌ม਺͝ͱʹઢܗճؼ/LOESS
    Feature Importance
    (Model Agnostic)
    

    View Slide

20. ղऍՄೳੑ
    .PEFM4QFDJpD .PEFM"HOPTUJD
    (MPCBM
    *OUFSQSFUBCJMJUZ
    w 3FHSFTTJPO$PF⒏DJFOUT
    w 'FBUVSF*NQPSUBODF
    ʜ
    w 4VSSPHBUF.PEFMT
    w 4FOTJUJWJUZ"OBMZTJT
    ʜ
    -PDBM
    *OUFSQSFUBCJMJUZ
    w .BYJNVN"DUJWBUJPO"OBMZTJT
    ʜ
    w -*.&
    w -0$0
    w 4)"1
    ʜ
    

    View Slide

21. Surrogate Models
    • ෳࡶͳϞσϧΛ୅ସ͢ΔγϯϓϧͳϞσϧΛ
    ௐ΂Δ
    Ideas on interpreting machine learning
    https://www.oreilly.com/ideas/ideas-on-interpreting-machine-learning
    Figure 14. An illustration of surrogate models for explaining a complex neural network. Figure courtesy of Patrick Hall and the H2O.ai team.
    

    View Slide

22. Sensitivity Analysis
    • ͋Δಛ௃ྔΛมԽͤͨ࣌͞ͷϞσϧͷग़ྗΛ
    ௐ΂Δ
    Sensitivity analysis for neural networks
    https://beckmw.wordpress.com/2013/10/07/sensitivity-analysis-for-neural-networks/
    

    View Slide

23. Partial Dependence
    • ͋Δಛ௃ྔΛมԽͤͨ࣌͞ͷϞσϧͷग़ྗΛ
    ௐ΂Δ
    library(pdp)
    library(ggplot2)
    pd <- partial(ca.fit,
    ɹɹpred.var = c("V11", “V9"),
    ɹchull = TRUE)
    autoplot(pd, contour = TRUE)
    

    View Slide

24. ղऍՄೳੑ
    .PEFM4QFDJpD .PEFM"HOPTUJD
    (MPCBM
    *OUFSQSFUBCJMJUZ
    w 3FHSFTTJPO$PF⒏DJFOUT
    w 'FBUVSF*NQPSUBODF
    ʜ
    w 1BSUJBM%FQFOEFODF
    w 4VSSPHBUF.PEFMT
    w 4FOTJUJWJUZ"OBMZTJT
    ʜ
    -PDBM
    *OUFSQSFUBCJMJUZ
    w .BYJNVN"DUJWBUJPO"OBMZTJT
    ʜ
    w -*.&
    w -0$0
    w 4)"1
    ʜ
    

    View Slide

25. LIMEͱ͸ʁ
    

    View Slide

26. Local Interpretable
    Model-agnostic Explanations
    

    View Slide

27. • ͋ΔσʔλͷۙลͰݩͷϞσϧΛ୅ସ͢ΔϞσϧΛ࡞੒ˠͦͷϞσϧͷ
    ಛ௃Λௐ΂Δ
    • Local Surrogate Model
    “Why Should I Trust You?” Explaining the Predictions of Any Classifier (Ribeiro, Singh, Guestrin
    2016)
    LIME
    

    View Slide

28. LIME
    • LIME͸ҎԼͷؔ਺Λ΋ͱʹσʔλ x ͷղऍΛಘΔ
    • G: ղऍ༻ͷֶशثͷू߹
    • L: ղऍֶ͍ͨ͠शثͱղऍ༻ͷֶशثͷ ΠxͷݩͰͷࠩ
    • f: ղऍֶ͍ͨ͠शث
    • Πx: σʔλ x ͱͷྨࣅ౓
    • Ω: ղऍ༻ͷֶशثͷෳࡶ͞ʹର͢Δേଇ߲
    • ۩ମతखஈ͸υϝΠϯʹґଘ
    

    View Slide

29. ςʔϒϧσʔλɾ෼ྨ໰୊ͷྫ
    • σʔλ x ͷपลͰαϯϓϦϯά
    • طఆͰ5,000
    • αϯϓϦϯάํ๏͸ม਺ͷछྨʹґଘ
    • Exponential KernelͰॏΈ෇͚
    • ม਺બ୒
    • Forward/Backward, LARSͳͲ
    • RidgeճؼͳͲ
    ˞1ZUIPO࣮૷ ޙड़
    ʹ΋ͱͮ͘
    

    View Slide

30. ύοέʔδ
    • Python
    • ࿦จஶऀ࡞
    • https://github.com/marcotcr/lime
    • R
    • ্هͷϙʔςΟϯά
    • https://github.com/thomasp85/lime
    install.packages(‘lime’)
    

    View Slide

31. LIME (R)
    • αϯϓϧ
    library(caret)
    library(lime)
    model <- train(iris[-5], iris[[5]], method = 'rf')
    explainer <- lime(iris[-5], model)
    explanations <- explain(iris[1, -5], explainer, n_labels = 1,
    n_features = 2)
    explanations
    model_type case label label_prob model_r2 model_intercept
    1 classification 1 setosa 1 0.3776584 0.2544468
    2 classification 1 setosa 1 0.3776584 0.2544468
    model_prediction feature feature_value feature_weight feature_desc
    1 0.7113922 Sepal.Width 3.5 0.02101138 3.3 < Sepal.Width
    2 0.7113922 Petal.Length 1.4 0.43593404 Petal.Length <= 1.60
    data prediction
    1 5.1, 3.5, 1.4, 0.2 1, 0, 0
    2 5.1, 3.5, 1.4, 0.2 1, 0, 0
    ֶशثΛ܇࿅
    ղऍ༻ͷΫϥεΛ࡞੒
    ղऍΛग़ྗ
    

    View Slide

32. LIME (R)
    model_type case label label_prob model_r2 model_intercept
    1 classification 1 setosa 1 0.3776584 0.2544468
    2 classification 1 setosa 1 0.3776584 0.2544468
    model_prediction feature feature_value feature_weight feature_desc
    1 0.7113922 Sepal.Width 3.5 0.02101138 3.3 < Sepal.Width
    2 0.7113922 Petal.Length 1.4 0.43593404 Petal.Length <= 1.60
    &YQMBOBUJPO
    NPEFM@UZQF 5IFUZQFPGUIFNPEFMVTFEGPSQSFEJDUJPO
    DBTF 5IFDBTFCFJOHFYQMBJOFE UIFSPXOBNFJODBTFT
    
    NPEFM@S 5IFRVBMJUZPGUIFNPEFMVTFEGPSUIFFYQMBOBUJPO
    NPEFM@JOUFSDFQU 5IFJOUFSDFQUPGUIFNPEFMVTFEGPSUIFFYQMBOBUJPO
    NPEFM@QSFEJDUJPO 5IFQSFEJDUJPOPGUIFPCTFSWBUJPOCBTFEPOUIFNPEFMVTFEGPSUIFFYQMBOBUJPO
    GFBUVSF 5IFGFBUVSFVTFEGPSUIFFYQMBOBUJPO
    GFBUVSF@WBMVF 5IFWBMVFPGUIFGFBUVSFVTFE
    GFBUVSF@XFJHIU 5IFXFJHIUPGUIFGFBUVSFJOUIFFYQMBOBUJPO
    GFBUVSF@EFTD "IVNBOSFBEBCMFEFTDSJQUJPOPGUIFGFBUVSFJNQPSUBODF
    

    View Slide

33. LIME (R)
    plot_features(explanations)
    ղऍΛϓϩοτ
    

    View Slide

34. ςΩετσʔλͷྫ
    • ϥϯμϜʹ୯ޠΛআ֎ͨ͠σʔλΛੜ੒
    • ίαΠϯྨࣅ౓ΛExponential KernelͰॏΈ෇͚
    • Ҏ߱ɺςʔϒϧσʔλͱಉ༷
    LIME - Local Interpretable Model-Agnostic Explanations
    https://homes.cs.washington.edu/~marcotcr/blog/lime/
    ˞1ZUIPO࣮૷ʹ΋ͱͮ͘
    

    View Slide

35. ը૾σʔλͷྫ
    Ideas on interpreting machine learning
    https://www.oreilly.com/ideas/ideas-on-interpreting-machine-learning
    Figure 15. An illustration of the LIME process in which a weighted linear model is used to explain a single prediction from a complex neural network. Figure courtesy of
    Marco Tulio Ribeiro; image used with permission.
    ˞1ZUIPO࣮૷ʹ΋ͱͮ͘
    • ϥϯμϜʹηάϝϯτԽͨ͠σʔλΛੜ੒
    • skimage.segmentationͷख๏͕ར༻Մ (quickshiftͳͲ)
    • ίαΠϯྨࣅ౓ͰॏΈ෇͚
    • Ҏ߱ɺςʔϒϧσʔλͱಉ༷
    ※RͰ͸ະ࣮૷
    

    View Slide

36. ·ͱΊ
    .PEFM4QFDJpD .PEFM"HOPTUJD
    (MPCBM
    *OUFSQSFUBCJMJUZ
    w 3FHSFTTJPO$PF⒏DJFOUT
    w 'FBUVSF*NQPSUBODF
    ʜ
    w 4VSSPHBUF.PEFMT
    w 4FOTJUJWJUZ"OBMZTJT
    ʜ
    -PDBM
    *OUFSQSFUBCJMJUZ
    w .BYJNVN"DUJWBUJPO"OBMZTJT
    ʜ
    w -*.&
    w -0$0
    w 4)"1
    ʜ
    

    View Slide

37. Enjoy!
    

    View Slide