Ranger中的SHAP重要性

Question

如果有一个二元分类问题：如何获取Ranger模型变量的Shap贡献？

样本数据：

library(ranger)
library(tidyverse)

# Binary Dataset
df <- iris
df$Target <- if_else(df$Species == "setosa",1,0)
df$Species <- NULL

# Train Ranger Model
model <- ranger(
  x = df %>%  select(-Target),
  y = df %>%  pull(Target))

我尝试了几个库(DALEX, shapr, fastshap, shapper)，但我没有得到任何解决方案。

我希望能像SHAPforxgboost一样获得一些结果：

• shap.values的输出，即变量的shap贡献
• shap.plot.summary

- PeCaDe

Answer 1

早上好！根据我所找到的信息，您可以按照以下方式使用ranger（）和fastshap()：

library(fastshap)
library(ranger)
library(tidyverse)
data(iris)
# Binary Dataset
df <- iris
df$Target <- if_else(df$Species == "setosa",1,0)
df$Species <- NULL
x <- df %>%  select(-Target)
# Train Ranger Model
model <- ranger(
  x = df %>%  select(-Target),
  y = df %>%  pull(Target))
# Prediction wrapper
pfun <- function(object, newdata) {
  predict(object, data = newdata)$predictions
}

# Compute fast (approximate) Shapley values using 10 Monte Carlo repetitions
system.time({  # estimate run time
  set.seed(5038)
  shap <- fastshap::explain(model, X = x, pred_wrapper = pfun, nsim = 10)
})

# Load required packages
library(ggplot2)
theme_set(theme_bw())

# Aggregate Shapley values
shap_imp <- data.frame(
  Variable = names(shap),
  Importance = apply(shap, MARGIN = 2, FUN = function(x) sum(abs(x)))
)

例如，对于变量重要性，您可以进行以下操作：

# Plot Shap-based variable importance
ggplot(shap_imp, aes(reorder(Variable, Importance), Importance)) +
  geom_col() +
  coord_flip() +
  xlab("") +
  ylab("mean(|Shapley value|)")

如果你想获得个别预测，可以采用以下方式：

# Plot individual explanations
expl <- fastshap::explain(model, X = x ,pred_wrapper = pfun, nsim = 10, newdata = x[1L, ])
autoplot(expl, type = "contribution")

所有这些信息都可以在这里找到，还有更多的内容：https://bgreenwell.github.io/fastshap/articles/fastshap.html 点击链接查看并解决您的疑惑！ :)

- Carles

Answer 2

我发布了两个R包来执行这些任务：一个是“kernelshap”（数据处理），另一个是“shapviz”（数据可视化）。

library(randomForest)
library(kernelshap)
Ilibrary(shapviz)

set.seed(1)
fit <- randomForest(Sepal.Length ~ ., data = iris,)

# bg_X is usually a small (50-200 rows) subset of the data

# Step 1: Calculate Kernel SHAP values
s <- kernelshap(fit, iris[-1], bg_X = iris)

# Step 2: Turn them into a shapviz object
sv <- shapviz(s)

# Step 3: Gain insights...
sv_importance(sv, show_numbers = TRUE)
sv_dependence(sv, v = "Petal.Length", color_var = "auto")

- Michael M