R 中的树模型机器学习
Sandro Raabe
Data Scientist
parsnip 决策树的超参数:
min_n:分裂节点所需的最小样本数tree_depth:树的最大深度cost_complexity:对树复杂度的惩罚parsnip 的默认值:
decision_tree(min_n = 20, tree_depth = 30, cost_complexity = 0.01)
调参的目标是找到超参数的最优组合。




spec_untuned <- decision_tree(min_n = tune(), tree_depth = tune()) %>% set_engine("rpart") %>% set_mode("classification")
Decision Tree Model Specification (classification)Main Arguments: tree_depth = tune() min_n = tune()
tune() 标记待调参的参数tree_grid <- grid_regular(parameters(spec_untuned),levels = 3 )
# A tibble: 9 x 2
min_n tree_depth
1 2 1
2 21 1
3 40 1
4 2 8
5 21 8
6 40 8
7 2 15
8 21 15
9 40 15
parameters()levels:每个超参数的网格点数
用法与参数:
metric_set() 包装的指标列表tune_results <- tune_grid(spec_untuned,outcome ~ .,resamples = my_folds,grid = tree_grid,metrics = metric_set(accuracy))
autoplot(tune_results)

# 选择表现最佳的参数 final_params <- select_best(tune_results)final_params
# A tibble: 1 x 3
min_n tree_depth .config
<int> <int> <chr>
1 2 8 Model4
# 填入规范中 best_spec <- finalize_model(spec_untuned, final_params)best_spec
Decision Tree Model Specification
(classification)
Main Arguments:
tree_depth = 8
min_n = 2
Computational engine: rpart
R 中的树模型机器学习