自適應重抽樣

R 的超參數調校

Dr. Shirin Elsinghorst

Senior Data Scientist

什麼是自適應重抽樣？

Grid Search 網格搜尋

計算所有超參數組合。

Random Search 隨機搜尋

隨機計算超參數組合的子集。

→ 最佳組合的評估在最後進行。

自適應重抽樣（Adaptive Resampling）

針對表現佳的組合附近，再重抽樣更多超參數組合。
因此自適應重抽樣更快、更有效率！

「Futility Analysis in the Cross-Validation of Machine Learning Models.」Max Kuhn；ARXIV 2014

caret 中的自適應重抽樣

trainControl: method = "adaptive_cv" + search = "random" + adaptive =

min：每個超參數的最少重抽樣次數
alpha：用於移除超參數的信心水準
method：線性模型用 "gls"，Bradley-Terry 用 "BT"
complete：若為 TRUE，則產生完整重抽樣集合

fitControl <- trainControl(method = "adaptive_cv",
                             adaptive = list(min = 2, alpha = 0.05, 
                                             method = "gls", complete = TRUE),
                             search = "random")

trainControl() + tuneLength = x

fitControl <- trainControl(method = "adaptive_cv", number = 3, repeats = 3,
                           adaptive = list(min = 2, 
                                           alpha = 0.05, 
                                           method = "gls", 
                                           complete = TRUE),
                             search = "random")

tic()
set.seed(42)
gbm_model_voters_adaptive <- train(turnout16_2016 ~ ., 
                                   data = voters_train_data, 
                                   method = "gbm", 
                                   trControl = fitControl,
                                   verbose = FALSE,
                                   tuneLength = 7)
toc()

共花費 1239.837 秒

自適應重抽樣

gbm_model_voters_adaptive

...
Resampling results across tuning parameters:
  shrinkage   interaction.depth  n.minobsinnode  n.trees  Accuracy   Kappa       Resamples
  0.07137493   5                  6              4152     0.9564654  0.02856571  9        
  0.08408739   5                 14               674     0.9547185  0.02098853  4        
  0.28552325   8                 15              3209     0.9568141  0.03024238  3        
  0.33663932  10                 13              2595     0.9571130  0.04250979  9        
  0.54251480   3                 24              3683     0.9482171  0.03568586  2        
  0.56406870   7                 25              4685     0.9549898  0.05284333  5        
  0.58695763   8                 24              1431     0.9520286  0.02742592  2        
Accuracy was used to select the optimal model using the largest value.
The final values used for the model were n.trees = 2595,
interaction.depth = 10, shrinkage = 0.3366393 and n.minobsinnode = 13.

開始動手寫程式吧！

R 的超參數調校