特徴量の重要度とモデルの解釈

モデルの解釈が重要な理由

# The interpretation toolkit we will use:
# - xgb.importance()      XGBoost global importance
# - xgb.plot.importance() Visualise XGBoost importance
# - ranger importance     Random forest importance
# - shapviz package       SHAP value computation + plots

cat('Packages needed: xgboost, ranger, shapviz')

xgb.importance() — 3つの指標

library(xgboost)
library(MASS)

X <- as.matrix(Boston[, -14])
y <- Boston[, 14]
dtrain <- xgb.DMatrix(data = X, label = y)

params <- list(objective = 'reg:squarederror', eta = 0.1,
               max_depth = 5, subsample = 0.8)
model  <- xgboost(data = dtrain, params = params,
                  nrounds = 100, verbose = 0)

imp <- xgb.importance(model = model, feature_names = colnames(X))
print(head(imp))

xgb.plot.importance()

imp <- xgb.importance(model = model, feature_names = colnames(X))

# Plot top 10 features by Gain
xgb.plot.importance(
  importance_matrix = imp,
  top_n  = 10,
  measure = 'Gain',
  main   = 'XGBoost Feature Importance (Gain)'
)

rangerの変数重要度

library(ranger)

# Impurity importance (fast)
rf_imp <- ranger(
  medv ~ ., data = MASS::Boston,
  num.trees  = 500,
  importance = 'impurity'
)
imp_impurity <- rf_imp$variable.importance

# Permutation importance (more reliable, slower)
rf_perm <- ranger(
  medv ~ ., data = MASS::Boston,
  num.trees  = 500,
  importance = 'permutation'
)
imp_permutation <- rf_perm$variable.importance

cbind(impurity = sort(imp_impurity, dec=TRUE),
      permutation = sort(imp_permutation, dec=TRUE))

rangerとXGBoostの重要度を比較する

# Align rankings for comparison
xgb_rank <- imp$Feature
rf_rank  <- names(sort(rf_perm$variable.importance, dec=TRUE))

# Spearman rank correlation between the two rankings
shared <- intersect(xgb_rank, rf_rank)
xgb_pos <- match(shared, xgb_rank)
rf_pos  <- match(shared, rf_rank)

cor(xgb_pos, rf_pos, method = 'spearman')

SHAP値とは

# Compute SHAP values natively from XGBoost
shap_matrix <- predict(
  model,
  newdata     = dtrain,
  predcontrib = TRUE  # returns SHAP contributions
)

# Result: matrix with one column per feature + BIAS column
dim(shap_matrix)  # rows x (features + 1)
colnames(shap_matrix)

shapvizパッケージの基本

library(shapviz)

# Build shapviz object from XGBoost model
shp <- shapviz(model, X_pred = X)

# Global importance: mean |SHAP| per feature
sv_importance(shp, kind = 'bar')

# Beeswarm plot (SHAP summary plot)
sv_importance(shp, kind = 'beeswarm')

ウォーターフォールプロット — ローカルな説明

library(shapviz)

shp <- shapviz(model, X_pred = X)

# Explain the prediction for observation 1
sv_waterfall(shp, row_id = 1) +
  ggplot2::labs(
    title = 'SHAP Waterfall — Observation 1',
    subtitle = 'How each feature contributed to this prediction'
  )

部分依存プロット

library(pdp)

# Partial dependence for 'lstat' in the Boston housing model
# We need a predict function wrapper for ranger
rf <- ranger(medv ~ ., data = MASS::Boston, num.trees = 300)

pd <- partial(
  rf,
  pred.var = 'lstat',
  train    = MASS::Boston,
  type     = 'regression'
)

plot(pd, type = 'l', lwd = 2,
     xlab = 'lstat', ylab = 'Partial Dependence',
     main = 'PDP: lstat vs medv')

SHAP依存プロット

library(shapviz)

shp <- shapviz(model, X_pred = X)

# SHAP dependence plot for 'lstat', coloured by 'rm'
sv_dependence(
  shp,
  v       = 'lstat',   # main feature on x-axis
  color_var = 'rm'     # interaction feature for colour
)

vipパッケージ — 統一された重要度

library(vip)

# Works with parsnip fitted models
lm_fit <- fit(
  linear_reg() |> set_engine('lm'),
  medv ~ ., data = MASS::Boston
)

vip(lm_fit, num_features = 10)

# Also works with ranger directly
vip(rf_imp, num_features = 10)

# Return importance as a data frame
vi(rf_imp)

確認問題

解釈のまとめ

library(shapviz)

# Complete interpretation workflow
shp <- shapviz(xgb_model, X_pred = X_matrix)

# 1. Global importance
sv_importance(shp, kind = 'beeswarm')

# 2. Local explanation for one observation
sv_waterfall(shp, row_id = 42)

# 3. Feature relationship
sv_dependence(shp, v = 'most_important_feature')