如何在R中基于训练集缩放测试集的age特征以避免信息泄漏？

我正在构建GBM分类器以预测某一目标变量。我的数据包含多个连续变量，仅需使用scale函数对其中的age特征进行缩放。为避免信息泄漏，需先缩放训练集的age，再基于训练集的规则缩放测试集的age，但我当前是分别缩放训练集和测试集的age，方法有误。现咨询如何在R（使用caret包）中正确实现，以下是我的现有代码：

for (i in (1:10)) {
  print(i)
  set.seed(i)
  
  IND = createDataPartition(y = MYData$Target_feature, p=0.8, list = FALSE)
  TRAIN_set = MYData[IND, ]
  TEST_set = MYData[-IND,]
  
  TRAIN_set$age = scale(TRAIN_set$age)
  TEST_set$age = scale(TEST_set$age)
  
  GBMModel <- train(Target_feature~., data = TRAIN_set,
                    method = "gbm",
                    metric="ROC",
                    trControl = ctrlCV,
                    tuneGrid = gbmGRID,
                    verbose = FALSE
                    
  )
  
  
  AUCs_Trn[i] = auc(roc(TRAIN_set$Target_feature,predict(GBMModel,TRAIN_set, type='prob')[,1]))
  AUCs_Tst[i] = auc(roc(TEST_set$Target_feature,predict(GBMModel,TEST_set, type='prob')[,1]))
  
}

注：我仅需缩放age特征。

核心是复用训练集缩放时的均值和标准差来处理测试集，避免测试集信息提前流入训练流程。具体修改如下：

for (i in (1:10)) {
  print(i)
  set.seed(i)
  
  IND = createDataPartition(y = MYData$Target_feature, p=0.8, list = FALSE)
  TRAIN_set = MYData[IND, ]
  TEST_set = MYData[-IND,]
  
  # 对训练集age缩放并保存缩放规则（均值、标准差）
  scale_obj <- scale(TRAIN_set$age)
  TRAIN_set$age <- scale_obj
  
  # 用训练集的缩放规则处理测试集age
  TEST_set$age <- scale(TEST_set$age, 
                        center = attr(scale_obj, "scaled:center"), 
                        scale = attr(scale_obj, "scaled:scale"))
  
  GBMModel <- train(Target_feature~., data = TRAIN_set,
                    method = "gbm",
                    metric="ROC",
                    trControl = ctrlCV,
                    tuneGrid = gbmGRID,
                    verbose = FALSE
                    
  )
  
  
  AUCs_Trn[i] = auc(roc(TRAIN_set$Target_feature,predict(GBMModel,TRAIN_set, type='prob')[,1]))
  AUCs_Tst[i] = auc(roc(TEST_set$Target_feature,predict(GBMModel,TEST_set, type='prob')[,1]))
  
}

关键说明 #

• scale()函数返回的对象附带scaled:center（训练集age的均值）和scaled:scale（训练集age的标准差）属性
• 测试集的age必须用训练集的均值和标准差做标准化，这样才能保证缩放逻辑完全基于训练数据，杜绝信息泄漏
• 这种方式仅对age特征做缩放，符合需求，其他特征不受影响

内容的提问来源于stack exchange，提问作者Programming Noob