🏆 アンサンブル学習の極意 - stacks

# アンサンブル学習エコシステムの読み込み library(tidymodels) library(stacks) library(finetune) library(tidyverse) # データ準備 - Boston住宅価格データセット data("Boston", package = "MASS") boston_data <- Boston %>% as_tibble() %>% mutate(medv_log = log(medv)) %>% select(-medv) # データ分割 set.seed(123) boston_split <- initial_split(boston_data, prop = 0.8, strata = medv_log) boston_train <- training(boston_split) boston_test <- testing(boston_split) # 交差検証設定 boston_folds <- vfold_cv(boston_train, v = 5, strata = medv_log)

# 前処理レシピ boston_recipe <- recipe(medv_log ~ ., data = boston_train) %>% step_normalize(all_numeric_predictors()) %>% step_corr(all_numeric_predictors(), threshold = 0.9) %>% step_zv(all_predictors()) # 1. 線形回帰 (Elastic Net) linear_spec <- linear_reg(penalty = tune(), mixture = tune()) %>% set_engine("glmnet") %>% set_mode("regression") # 2. Random Forest rf_spec <- rand_forest( trees = 1000, mtry = tune(), min_n = tune() ) %>% set_engine("ranger") %>% set_mode("regression") # 3. XGBoost xgb_spec <- boost_tree( trees = tune(), tree_depth = tune(), learn_rate = tune() ) %>% set_engine("xgboost") %>% set_mode("regression")

# スタッキング用制御オプション control_stack <- control_stack_grid() # 線形回帰のチューニング linear_res <- workflow() %>% add_recipe(boston_recipe) %>% add_model(linear_spec) %>% tune_grid( resamples = boston_folds, grid = 15, control = control_stack, metrics = metric_set(rmse, mae, rsq) ) # Random Forestのチューニング rf_res <- workflow() %>% add_recipe(boston_recipe) %>% add_model(rf_spec) %>% tune_grid( resamples = boston_folds, grid = 12, control = control_stack, metrics = metric_set(rmse, mae, rsq) ) # XGBoostのチューニング xgb_res <- workflow() %>% add_recipe(boston_recipe) %>% add_model(xgb_spec) %>% tune_grid( resamples = boston_folds, grid = 15, control = control_stack, metrics = metric_set(rmse, mae, rsq) )

# スタッキングモデルの構築 boston_stack <- stacks() %>% add_candidates(linear_res) %>% add_candidates(rf_res) %>% add_candidates(xgb_res) cat("候補モデル数:", length(boston_stack), "\n") # メタラーナー（Lasso）でのブレンディング boston_blend <- boston_stack %>% blend_predictions( penalty = 10^seq(-4, -0.5, 0.25), mixture = 1 # Lasso回帰 ) # 最終アンサンブルの学習 boston_ensemble <- boston_blend %>% fit_members() # テストデータでの評価 ensemble_pred <- predict(boston_ensemble, boston_test) %>% bind_cols(boston_test) # 性能指標の計算 ensemble_metrics <- ensemble_pred %>% metrics(truth = medv_log, estimate = .pred)

# モデルの重みを確認 model_weights <- collect_parameters(boston_ensemble, "blend_predictions") %>% filter(coef != 0) %>% arrange(desc(abs(coef))) print(model_weights) # 予測の信頼性分析 prediction_reliability <- ensemble_pred %>% mutate( residual = medv_log - .pred, abs_residual = abs(residual), relative_error = abs_residual / medv_log * 100 ) %>% summarise( mean_residual = mean(residual), std_residual = sd(residual), mean_abs_error = mean(abs_residual), median_rel_error = median(relative_error), q90_rel_error = quantile(relative_error, 0.9) )

# 問題特性に応じたベースモデル選択戦略 # 🔢 数値データ中心の場合 numeric_focused_models <- list( # 線形関係を捉える linear = linear_reg(penalty = tune(), mixture = tune()), # 非線形関係を捉える spline = gen_additive_mod(select_features = tune()), # 複雑な相互作用 xgboost = boost_tree(trees = tune(), tree_depth = tune()) ) # 📊 カテゴリカルデータ多い場合 categorical_focused_models <- list( # カテゴリ処理に強い random_forest = rand_forest(mtry = tune(), trees = 1000), # カテゴリ埋め込み lightgbm = boost_tree(engine = "lightgbm"), # ナイーブベイズ naive_bayes = naive_Bayes(smoothness = tune()) ) # ⏰ 時系列データの場合 timeseries_focused_models <- list( # トレンド・季節性 prophet = prophet_reg(seasonality_yearly = tune()), # 自己回帰 arima = arima_reg(), # 機械学習アプローチ mars = mars(num_terms = tune(), prod_degree = tune()) )