1

我正在使用插入符号来拟合 gbm 模型。当我打电话时trainedGBM$finalModel$fit，我得到的输出看起来正确。

但是当我打电话时predict(trainedGBM$finalModel, origData, type="response")，我得到了非常不同的结果，predict(trainedGBM$finalModel, type="response")即使附加了 origData 也会产生不同的结果。在我看来，这些调用应该产生相同的输出。有人可以帮我确定问题吗？

library(caret)
library(gbm)

attach(origData)
gbmGrid <- expand.grid(.n.trees = c(2000), 
                       .interaction.depth = c(14:20), 
                       .shrinkage = c(0.005))
trainedGBM <- train(y ~ ., method = "gbm", distribution = "gaussian", 
                    data = origData, tuneGrid = gbmGrid, 
                    trControl = trainControl(method = "repeatedcv", number = 10, 
                                             repeats = 3, verboseIter = FALSE, 
                                             returnResamp = "all"))
ntrees <- gbm.perf(trainedGBM$finalModel, method="OOB")
data.frame(y, 
           finalModelFit = trainedGBM$finalModel$fit, 
           predictDataSpec = predict(trainedGBM$finalModel, origData, type="response", n.trees=ntrees), 
           predictNoDataSpec = predict(trainedGBM$finalModel, type="response", n.trees=ntrees))

上面的代码产生以下部分结果：

   y finalModelFit predictDataSpec predictNoDataSpec
9000     6138.8920        2387.182          2645.993
5000     3850.8817        2767.990          2467.157
3000     3533.1183        2753.551          2044.578
2500     1362.9802        2672.484          1972.361
1500     5080.2112        2449.185          2000.568
 750     2284.8188        2728.829          2063.829
1500     2672.0146        2359.566          2344.451
5000     3340.5828        2435.137          2093.939
   0     1303.9898        2377.770          2041.871
 500      879.9798        2691.886          2034.307
3000     2928.4573        2327.627          1908.876
4

1 回答 1

7

根据您的gbmGrid，只有您的交互深度会在 14 到 20 之间变化，并且树的收缩和数量分别固定为 0.005 和 2000。其TrainedGBM设计仅发现当前的最佳交互水平。然后，您的ntrees计算gbm.perf结果是，鉴于最佳交互水平在 14 到 20 之间，基于 OOB 标准的最佳树数是多少。因为预测取决于模型中树的数量，所以将使用基于训练的 GBM 的ntrees = 2000预测，而基于 的预测将使用从该函数估计gbm.perf的最佳数量 。这将解释您的和 ntrees之间的差异。trainedGBM$finalModel$fitpredict(trainedGBM$finalModel, type="response", n.trees=ntrees)

展示一个基于 iris 数据集的示例，使用 gbm 作为分类而不是回归模型

library(caret)
library(gbm)

set.seed(42)

gbmGrid <- expand.grid(.n.trees = 100, 
                   .interaction.depth = 1:4, 
                   .shrinkage = 0.05)


trainedGBM <- train(Species ~ ., method = "gbm", distribution='multinomial',
                data = iris, tuneGrid = gbmGrid, 
                trControl = trainControl(method = "repeatedcv", number = 10, 
                                         repeats = 3, verboseIter = FALSE, 
                                         returnResamp = "all"))
print(trainedGBM)

给予

# Resampling results across tuning parameters:

#  interaction.depth  Accuracy  Kappa  Accuracy SD  Kappa SD
#   1                  0.947     0.92   0.0407       0.061   
#   2                  0.947     0.92   0.0407       0.061   
#   3                  0.944     0.917  0.0432       0.0648  
#   4                  0.944     0.917  0.0395       0.0592  

# Tuning parameter 'n.trees' was held constant at a value of 100
# Tuning parameter 'shrinkage' was held constant at a value of 0.05
# Accuracy was used to select the optimal model using  the largest value.
# The final values used for the model were interaction.depth = 1, n.trees = 100
# and shrinkage = 0.05.

以最佳交互深度为条件找到最佳树数：

ntrees <-  gbm.perf(trainedGBM$finalModel, method="OOB")
# Giving ntrees = 50

如果我们通过改变树的数量和交互深度来训练模型：

gbmGrid2 <- expand.grid(.n.trees = 1:100, 
                   .interaction.depth = 1:4, 
                   .shrinkage = 0.05)

trainedGBM2 <- train(Species ~ ., method = "gbm", 
                data = iris, tuneGrid = gbmGrid2, 
                trControl = trainControl(method = "repeatedcv", number = 10, 
                                         repeats = 3, verboseIter = FALSE, 
                                         returnResamp = "all"))

print(trainedGBM2) 

# Tuning parameter 'shrinkage' was held constant at a value of 0.05
# Accuracy was used to select the optimal model using  the largest value.
# The final values used for the model were interaction.depth = 2, n.trees = 39
# and shrinkage = 0.05.

请注意，当我们同时改变树的数量和交互深度时，树的最佳数量非常接近为 计算的值gbm.perf。

于 2013-07-14T02:39:27.603 回答