Merge pull request #4 from mayer79/update_shap2

Update SHAP markdown and html

14 - SHAP/README.md

@@ -17,4 +17,4 @@ The folder "rdata" contains the Parquet file "df.parquet", representing 1 mio ro
 ## Code output
 
 - Python: Click on the ipynb file
-- R: [shap_tutorial.html](https://htmlpreview.github.io/?https://github.com/JSchelldorfer/ActuarialDataScience/blob/master/14%20-%20SHAP/shap_tutorial.html) (please click on this link to preview html output; in order to see the table of contents properly, download the html and open it in your browser.)
+- R: [shap.html](https://htmlpreview.github.io/?https://github.com/JSchelldorfer/ActuarialDataScience/blob/master/docs/shap.html) (please click on this link to preview html output; in order to see the table of contents properly, download the html and open it in your browser.)

14 - SHAP/shap_tutorial.Rmd

@@ -34,16 +34,17 @@ library(farff)       # 1.1.1 (required by OpenML)
 library(OpenML)      # 1.12  (if you download the data from OpenML)
 
 library(withr)       # 2.5.0
-library(tidyverse)   # 1.3.2
+library(tidyverse)   # 2.0.0
 
 # Models
-library(caret)       # 6.0-93
-library(keras)       # 2.11.0
+library(caret)       # 6.0.94
+library(keras)       # 2.11.1
 library(lightgbm)    # 3.3.5
 
 # SHAP
-library(shapviz)     # 0.6.0
-library(kernelshap)  # 0.3.4
+library(shapviz)     # 0.9.1
+library(kernelshap)  # 0.3.7
+library(patchwork)   # 1.1.2
 
 # df <- arrow::read_parquet("rdata/df.parquet")
 df <- getOMLDataSet(data.id = 45106L)$data
@@ -230,12 +231,10 @@ sv_importance(shap_lgb, kind = "bee")
 Each figure uses the potentially strongest interacting feature on the color scale. Therefore, the colors differ from picture to picture (and later also from model to model).
 
 ```{r}
-for (xvar in x) {
-  (sv_dependence(shap_lgb, xvar, alpha = 0.5) + 
-    ggtitle(xvar) +
-    ylim(-0.5, 1.05)) %>% 
-  print()
-}
+theme_set(theme_gray(base_size = 8))
+
+sv_dependence(shap_lgb, x, alpha = 0.5) &
+  ylim(-0.5, 1.05)
 ```
 
 ## GLM with Kernel SHAP
@@ -247,12 +246,8 @@ system.time(
   shap_glm <- shapviz(kernelshap(fit_glm, X = X_explain, bg_X = bg, verbose = FALSE))
 )
 
-for (xvar in x) {
-  (sv_dependence(shap_glm, xvar, alpha = 0.5) + 
-    ggtitle(xvar) +
-    ylim(-0.5, 1.05)) %>% 
-  print()
-}
+sv_dependence(shap_glm, x, alpha = 0.5) &
+  ylim(-0.5, 1.05)
 ```
 
 ## Neural net with Kernel SHAP
@@ -273,12 +268,8 @@ system.time(
 )
 
 # Dependence plots
-for (xvar in x) {
-  (sv_dependence(shap_nn, xvar, alpha = 0.5) + 
-    ggtitle(xvar) +
-    ylim(-0.5, 1.05)) %>% 
-  print()
-}
+sv_dependence(shap_nn, x, alpha = 0.5) &
+  ylim(-0.5, 1.05)
 ```
 
 ## True model with Kernel SHAP
@@ -296,10 +287,6 @@ system.time(
   )
 )
 
-for (xvar in x) {
-  (sv_dependence(shap_truth, xvar, alpha = 0.5) + 
-    ggtitle(xvar) +
-    ylim(-0.5, 1.05)) %>% 
-  print()
-}
+sv_dependence(shap_truth, x, alpha = 0.5) &
+  ylim(-0.5, 1.05)
 ```