Model Agnostic Interpretability

“Interpretability”

Permutation Importance

R

Code

library(randomForest)

set.seed(12345)
rf.ames <- randomForest(Sale_Price ~ ., data = training.df, ntree = 250, importance = TRUE)

varImpPlot(rf.ames,
           sort = TRUE,
           n.var = 10,
           main = "Top 10 - Variable Importance", type = 1)

Code

importance(rf.ames)

                %IncMSE IncNodePurity
Bedroom_AbvGr 12.021934  1.934816e+11
Year_Built    49.013449  2.603400e+12
Mo_Sold        2.105040  1.851732e+11
Lot_Area      11.714265  5.719501e+11
Street         5.072419  4.383793e+09
Central_Air   14.485207  1.011408e+11
First_Flr_SF  27.776232  1.870211e+12
Second_Flr_SF 21.568185  5.785342e+11
Full_Bath     11.247856  6.078854e+11
Half_Bath      9.059787  1.016659e+11
Fireplaces    19.593721  5.671476e+11
Garage_Area   22.819303  2.140693e+12
Gr_Liv_Area   25.768069  2.497100e+12
TotRms_AbvGrd 11.695212  4.433089e+11

Code

training.df$pred_rf <- predict(rf.ames)

Code

library(iml)

lm.ames <- lm(Sale_Price ~ ., data = training)

summary(lm.ames)

Call:
lm(formula = Sale_Price ~ ., data = training)

Residuals:
    Min      1Q  Median      3Q     Max 
-505730  -19610   -1560   17462  254027 

Coefficients:
                Estimate Std. Error t value Pr(>|t|)    
(Intercept)   -1.513e+06  8.174e+04 -18.514  < 2e-16 ***
Bedroom_AbvGr -1.251e+04  1.512e+03  -8.270 2.39e-16 ***
Year_Built     7.700e+02  4.226e+01  18.218  < 2e-16 ***
Mo_Sold       -5.963e+02  3.226e+02  -1.849  0.06467 .  
Lot_Area       3.471e-01  1.167e-01   2.975  0.00297 ** 
StreetPave     2.173e+04  1.525e+04   1.425  0.15425    
Central_AirY   6.043e+03  3.875e+03   1.560  0.11900    
First_Flr_SF   9.762e+01  2.059e+01   4.742 2.26e-06 ***
Second_Flr_SF  7.404e+01  2.059e+01   3.596  0.00033 ***
Full_Bath     -3.270e+03  2.443e+03  -1.339  0.18077    
Half_Bath     -6.372e+03  2.516e+03  -2.533  0.01139 *  
Fireplaces     1.056e+04  1.576e+03   6.704 2.61e-11 ***
Garage_Area    5.816e+01  5.404e+00  10.762  < 2e-16 ***
Gr_Liv_Area    1.601e+01  2.033e+01   0.787  0.43115    
TotRms_AbvGrd -5.310e+02  1.087e+03  -0.488  0.62539    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 39360 on 2036 degrees of freedom
Multiple R-squared:  0.7526,    Adjusted R-squared:  0.7509 
F-statistic: 442.3 on 14 and 2036 DF,  p-value: < 2.2e-16

Code

linear_pred <- Predictor$new(lm.ames, data = training[,-1], 
                             y = training$Sale_Price, type = "response")
plot(FeatureImp$new(linear_pred, loss = "mse"))

Python

Individual Conditional Expectation (ICE)

R

Code

library(patchwork)

set.seed(12345)
forest_pred <- Predictor$new(rf.ames, data = training[,-1], 
                             y = training$Sale_Price, type = "response")

ice_plot <- FeatureEffects$new(forest_pred, method = "ice")
ice_plot$plot(c("Garage_Area"))

Code

ice_plot$plot()

Python

Partial Dependence

R

Code

pd_plot <- FeatureEffects$new(forest_pred, method = "pdp")
pd_plot$plot(c("Garage_Area"))

Code

pd_plot$plot()

Warning: Removed 2 rows containing missing values (`geom_col()`).
Removed 2 rows containing missing values (`geom_col()`).

Code

pdice_plot <- FeatureEffects$new(forest_pred, method = "pdp+ice")
pdice_plot$plot(c("Garage_Area"))

Code

pdice_plot$plot()

Code

pdice_plot$plot(c("Mo_Sold"))

Python

Accumulated Local Effects (ALE)

R

Code

ale_plot <- FeatureEffects$new(forest_pred, method = "ale")
ale_plot$plot(c("Garage_Area"))

Code

ale_plot$plot()

Python

Local Interpretable Model-Agnostic Explanations (LIME)

R

Code

point <- 1328
lime.explain <- LocalModel$new(forest_pred, x.interest = training[point,-1], k = 5)

Loading required package: glmnet

Loaded glmnet 4.1-7

Loading required package: gower

Code

plot(lime.explain) + theme_bw()

Code

point <- 1000
lime.explain <- LocalModel$new(forest_pred, x.interest = training[point,-1], k = 5)
plot(lime.explain)

Python

Shapley Values

R

Code

point <- 1328
shap <- Shapley$new(forest_pred, x.interest = training[point,-1])
shap$plot()

Code

point <- 1000
shap <- Shapley$new(forest_pred, x.interest = training[point,-1])
shap$plot()

Python