Dec 4th, 2017
怎麼挑模型
Model Selection 根據問題與資料型態
Cross Validation 交叉驗證
黑魔法！
解釋
Explainable AI
Explainable AI
確認它的判斷合理（verification of the system）
改良它的算法（improvement of the system）
Explainable AI
從它身上學習（learning from the system）
- AlphaGO
符合法規要求（compliance to legislation）
- 歐盟 GDPR 規定使用者有「要求解釋的權力」（right to explanation）
LIME
Local Interpretable Model-Agnostic Explanations
LIME
Local Interpretable Model-Agnostic Explanations
LIME
- Local Interpretable Model-Agnostic Explanations
LIME
- Local Interpretable Model-Agnostic Explanations
LIME
- Local Interpretable Model-Agnostic Explanations
LIME
- Local Interpretable Model-Agnostic Explanations
LIME
KDD2016 paper 573
LIME
LIME
基本概念
- For each prediction to explain, permute the observation n times.
- Let the complex model predict the outcome of all permuted observations.
- Calculate the distance from all permutations to the original observation.
- Convert the distance to a similarity score.
LIME
基本概念
- Select m features best describing the complex model outcome from the permuted data.
- Fit a simple model to the permuted data, explaining the complex model outcome with the m features from the permuted data weighted by its similarity to the original observation.
- Extract the feature weights from the simple model and use these as explanations for the complex models local behavior.
LIME
Feature Selection
none: Use all features for the explanation.forward selection: Features are added one by one based on their improvements to a ridge regression fit of the complex model outcome.highest weights: The m features with highest absolute weight in a ridge regression fit of the complex model outcome are chosen.
LIME
Feature Selection
lasso: The m features that are least prone to shrinkage based on the regularization path of a lasso fit of the complex model outcome is chosen.tree: A tree is fitted with log2(m) splits, to use at max m features. It may possibly select less.auto: Usesforward selectionif m <= 6 and otherwisehighest weights.
Demo
Demo
library(lime)
library(MASS)
data(biopsy)
# First we'll clean up the data a bit
biopsy$ID <- NULL
biopsy <- na.omit(biopsy)
names(biopsy) <- c('clump thickness', 'uniformity of cell size',
'uniformity of cell shape', 'marginal adhesion',
'single epithelial cell size', 'bare nuclei',
'bland chromatin', 'normal nucleoli', 'mitoses',
'class')
Demo
# Now we'll fit a linear discriminant model on all but 4 cases set.seed(4) test_set <- sample(seq_len(nrow(biopsy)), 100) prediction <- biopsy$class biopsy$class <- NULL model <- lda(biopsy[-test_set, ], prediction[-test_set]) sum(predict(model, biopsy[test_set, ])$class == prediction[test_set])/100
## [1] 0.96
Demo
# Train the explainer explainer <- lime(biopsy[-test_set,], model, bin_continuous = TRUE, quantile_bins = FALSE) # Use the explainer on new observations explanation <- explain(biopsy[test_set[1:4], ], explainer, n_labels = 1, n_features = 4) tibble::glimpse(explanation)
## Observations: 16 ## Variables: 13 ## $ model_type <chr> "classification", "classification", "classifi... ## $ case <chr> "416", "416", "416", "416", "7", "7", "7", "7... ## $ label <chr> "benign", "benign", "benign", "benign", "beni... ## $ label_prob <dbl> 0.9964864, 0.9964864, 0.9964864, 0.9964864, 0... ## $ model_r2 <dbl> 0.5659044, 0.5659044, 0.5659044, 0.5659044, 0... ## $ model_intercept <dbl> 0.08837631, 0.08837631, 0.08837631, 0.0883763... ## $ model_prediction <dbl> 1.0244738, 1.0244738, 1.0244738, 1.0244738, 0... ## $ feature <chr> "normal nucleoli", "bare nuclei", "uniformity... ## $ feature_value <int> 5, 3, 3, 3, 1, 10, 1, 1, 5, 10, 10, 3, 1, 1, ... ## $ feature_weight <dbl> -0.018041571, 0.573050022, 0.202345467, 0.178... ## $ feature_desc <chr> "3.25 < normal nucleoli <= 5.50", "bare nucle... ## $ data <list> [[3, 3, 2, 6, 3, 3, 3, 5, 1], [3, 3, 2, 6, 3... ## $ prediction <list> [[0.9964864, 0.003513577], [0.9964864, 0.003...
Demo
explanation <- explain(biopsy[test_set[1:4], ], explainer, n_labels = 1,
n_features = 4, kernel_width = 0.5, feature_select = "auto")
explanation[, 2:9]
## case label label_prob model_r2 model_intercept model_prediction ## 1 416 benign 0.9964864 0.4804993 0.4323631 1.0029394 ## 2 416 benign 0.9964864 0.4804993 0.4323631 1.0029394 ## 3 416 benign 0.9964864 0.4804993 0.4323631 1.0029394 ## 4 416 benign 0.9964864 0.4804993 0.4323631 1.0029394 ## 5 7 benign 0.9244742 0.4680113 0.3216358 0.6370384 ## 6 7 benign 0.9244742 0.4680113 0.3216358 0.6370384 ## 7 7 benign 0.9244742 0.4680113 0.3216358 0.6370384 ## 8 7 benign 0.9244742 0.4680113 0.3216358 0.6370384 ## 9 207 malignant 0.9999911 0.6543314 0.1583423 1.0001967 ## 10 207 malignant 0.9999911 0.6543314 0.1583423 1.0001967 ## 11 207 malignant 0.9999911 0.6543314 0.1583423 1.0001967 ## 12 207 malignant 0.9999911 0.6543314 0.1583423 1.0001967 ## 13 195 benign 0.9999981 0.4631399 0.5493353 1.0047116 ## 14 195 benign 0.9999981 0.4631399 0.5493353 1.0047116 ## 15 195 benign 0.9999981 0.4631399 0.5493353 1.0047116 ## 16 195 benign 0.9999981 0.4631399 0.5493353 1.0047116 ## feature feature_value ## 1 mitoses 1 ## 2 bare nuclei 3 ## 3 clump thickness 3 ## 4 uniformity of cell size 3 ## 5 mitoses 1 ## 6 bare nuclei 10 ## 7 clump thickness 1 ## 8 uniformity of cell size 1 ## 9 mitoses 1 ## 10 uniformity of cell size 10 ## 11 clump thickness 10 ## 12 uniformity of cell shape 9 ## 13 mitoses 1 ## 14 bare nuclei 1 ## 15 clump thickness 3 ## 16 uniformity of cell size 1
Demo
plot_features(explanation, ncol = 2)
