rpart 的结果是根,但数据显示信息增益

result of rpart is a root, but data shows Information Gain

我有一个事件率低于 3% 的数据集(即大约有 700 条记录 class 1 和 27000 条记录 class 0)。

ID          V1  V2      V3  V5      V6  Target
SDataID3    161 ONE     1   FOUR    0   0
SDataID4    11  TWO     2   THREE   2   1
SDataID5    32  TWO     2   FOUR    2   0
SDataID7    13  ONE     1   THREE   2   0
SDataID8    194 TWO     2   FOUR    0   0
SDataID10   63  THREE   3   FOUR    0   1
SDataID11   89  ONE     1   FOUR    0   0
SDataID13   78  TWO     2   FOUR    0   0
SDataID14   87  TWO     2   THREE   1   0
SDataID15   81  ONE     1   THREE   0   0
SDataID16   63  ONE     3   FOUR    0   0
SDataID17   198 ONE     3   THREE   0   0
SDataID18   9   TWO     3   THREE   0   0
SDataID19   196 ONE     2   THREE   2   0
SDataID20   189 TWO     2   ONE     1   0
SDataID21   116 THREE   3   TWO     0   0
SDataID24   104 ONE     1   FOUR    0   0
SDataID25   5   ONE     2   ONE     3   0
SDataID28   173 TWO     3   FOUR    0   0
SDataID29   5   ONE     3   ONE     3   0
SDataID31   87  ONE     3   FOUR    3   0
SDataID32   5   ONE     2   THREE   1   0
SDataID34   45  ONE     1   FOUR    0   0
SDataID35   19  TWO     2   THREE   0   0
SDataID37   133 TWO     2   FOUR    0   0
SDataID38   8   ONE     1   THREE   0   0
SDataID39   42  ONE     1   THREE   0   0
SDataID43   45  ONE     1   THREE   1   0
SDataID44   45  ONE     1   FOUR    0   0
SDataID45   176 ONE     1   FOUR    0   0
SDataID46   63  ONE     1   THREE   3   0

我正在尝试使用决策树找出拆分。但是树的结果只有 1 个根。

> library(rpart)
> tree <- rpart(Target ~ ., data=subset(train, select=c( -Record.ID) ),method="class")
> printcp(tree)

Classification tree:
rpart(formula = Target ~ ., data = subset(train, select = c(-Record.ID)), method = "class")

Variables actually used in tree construction:
character(0)

Root node error: 749/18239 = 0.041066

n= 18239 

  CP nsplit rel error xerror xstd
1  0      0         1      0    0

阅读 Whosebug 上的大部分资源后,我 loosened/tweaked 控制参数为我提供了所需的决策树。

> tree <- rpart(Target ~ ., data=subset(train, select=c( -Record.ID) ),method="class" ,control =rpart.control(minsplit = 1,minbucket=2, cp=0.00002))
> printcp(tree)

Classification tree:
rpart(formula = Target ~ ., data = subset(train, select = c(-Record.ID)), 
    method = "class", control = rpart.control(minsplit = 1, minbucket = 2, 
        cp = 2e-05))

Variables actually used in tree construction:
[1] V5         V2                     V1          
[4] V3         V6

Root node error: 749/18239 = 0.041066

n= 18239 

          CP nsplit rel error xerror     xstd
1 0.00024275      0   1.00000 1.0000 0.035781
2 0.00019073     20   0.99466 1.0267 0.036235
3 0.00016689     34   0.99199 1.0307 0.036302
4 0.00014835     54   0.98798 1.0334 0.036347
5 0.00002000     63   0.98665 1.0427 0.036504

当我修剪这棵树时,结果是一棵只有一个节点的树。

> pruned.tree <- prune(tree, cp = tree$cptable[which.min(tree$cptable[,"xerror"]),"CP"])
> printcp(pruned.tree)

Classification tree:
rpart(formula = Target ~ ., data = subset(train, select = c(-Record.ID)), 
    method = "class", control = rpart.control(minsplit = 1, minbucket = 2, 
        cp = 2e-05))

Variables actually used in tree construction:
character(0)

Root node error: 749/18239 = 0.041066

n= 18239 

          CP nsplit rel error xerror     xstd
1 0.00024275      0         1      1 0.035781

树不应该只给出根节点,因为从数学上讲,在给定节点(提供的示例)上我们正在获得信息增益。我不知道我是否在修剪时犯了错误,或者 rpart 在处理低事件率数据集时存在问题?

NODE    p       1-p     Entropy         Weights         Ent*Weight      # Obs
Node 1  0.032   0.968   0.204324671     0.351398601     0.071799404     10653
Node 2  0.05    0.95    0.286396957     0.648601399     0.185757467     19663

Sum(Ent*wght)       0.257556871 
Information gain    0.742443129 

您提供的数据未反映两个目标的比率 类,因此我调整了数据以更好地反映这一点(参见数据部分):

> prop.table(table(train$Target))

         0          1 
0.96707581 0.03292419 

> 700/27700
[1] 0.02527076

比率现在比较接近...

library(rpart)
tree <- rpart(Target ~ ., data=train, method="class")
printcp(tree)

结果:

Classification tree:
rpart(formula = Target ~ ., data = train, method = "class")

Variables actually used in tree construction:
character(0)

Root node error: 912/27700 = 0.032924

n= 27700 

  CP nsplit rel error xerror xstd
1  0      0         1      0    0

现在,您只看到第一个模型的根节点的原因可能是因为您的目标极度不平衡 类,因此您的自变量无法提供足够的信息种树的信息。我的样本数据有 3.3% 的事件率,但你的只有 2.5% 左右!

正如您提到的,有一种方法可以强制 rpart 种植树。那就是覆盖默认的复杂度参数(cp)。复杂性度量是树的大小和树分离目标的程度的组合 类。来自?rpart.control"Any split that does not decrease the overall lack of fit by a factor of cp is not attempted"。这意味着此时您的模型没有超出根节点的拆分,这足以降低复杂性级别以供 rpart 考虑。我们可以通过设置低值或负值 cp 来放宽被认为是 "enough" 的阈值(负值 cp 基本上会迫使树增长到最大尺寸)。

tree <- rpart(Target ~ ., data=train, method="class" ,parms = list(split = 'information'), 
              control =rpart.control(minsplit = 1,minbucket=2, cp=0.00002))
printcp(tree)

结果:

Classification tree:
rpart(formula = Target ~ ., data = train, method = "class", parms = list(split = "information"), 
    control = rpart.control(minsplit = 1, minbucket = 2, cp = 2e-05))

Variables actually used in tree construction:
[1] ID V1 V2 V3 V5 V6

Root node error: 912/27700 = 0.032924

n= 27700 

           CP nsplit rel error xerror     xstd
1  4.1118e-04      0   1.00000 1.0000 0.032564
2  3.6550e-04     30   0.98355 1.0285 0.033009
3  3.2489e-04     45   0.97807 1.0702 0.033647
4  3.1328e-04    106   0.95504 1.0877 0.033911
5  2.7412e-04    116   0.95175 1.1031 0.034141
6  2.5304e-04    132   0.94737 1.1217 0.034417
7  2.1930e-04    149   0.94298 1.1458 0.034771
8  1.9936e-04    159   0.94079 1.1502 0.034835
9  1.8275e-04    181   0.93640 1.1645 0.035041
10 1.6447e-04    193   0.93421 1.1864 0.035356
11 1.5664e-04    233   0.92654 1.1853 0.035341
12 1.3706e-04    320   0.91228 1.2083 0.035668
13 1.2183e-04    344   0.90899 1.2127 0.035730
14 9.9681e-05    353   0.90789 1.2237 0.035885
15 2.0000e-05    364   0.90680 1.2259 0.035915

如您所见,树的大小已将复杂度级别至少降低 cp。需要注意两点:

  1. 在零 nsplit 时,CP 已经低至 0.0004,而 rpart 中的默认值 cp 设置为 0.01。
  2. nsplit == 0 开始,交叉验证错误 (xerror) 随着拆分次数的增加而增加

这两个都表明您的模型在 nsplit == 0 及以后过度拟合数据,因为在您的模型中添加更多自变量并没有添加足够的信息(CP 减少不足)来减少交叉验证错误.话虽这么说,你的根节点模型在这种情况下最好的模型,这解释了为什么你的初始模型只有根节点。

pruned.tree <- prune(tree, cp = tree$cptable[which.min(tree$cptable[,"xerror"]),"CP"])
printcp(pruned.tree)

结果:

Classification tree:
rpart(formula = Target ~ ., data = train, method = "class", parms = list(split = "information"), 
    control = rpart.control(minsplit = 1, minbucket = 2, cp = 2e-05))

Variables actually used in tree construction:
character(0)

Root node error: 912/27700 = 0.032924

n= 27700 

          CP nsplit rel error xerror     xstd
1 0.00041118      0         1      1 0.032564

至于修剪部分,现在更清楚为什么你的修剪树是根节点树,因为超过 0 个分裂的树会增加交叉验证错误。采用具有最小 xerror 的树将使您得到预期的根节点树。

信息增益基本上告诉你每次拆分增加了多少"information"。所以从技术上讲,每个 拆分都有一定程度的信息增益,因为您在模型中添加了更多变量(信息增益始终是非负的)。您应该考虑的是,额外的收益(或没有收益)是否足以减少错误,使您能够保证更复杂的模型。因此,偏差和方差之间的权衡。

在这种情况下,减少 cp 并随后修剪生成的树对您来说没有任何意义。因为通过设置较低的 cp,你告诉 rpart 即使它过度拟合也要进行拆分,同时修剪 "cuts" 所有过度拟合的节点。

数据:

请注意,我正在为每一列和样本打乱行,而不是对行索引进行抽样。这是因为您提供的数据可能不是原始数据集的随机样本(可能有偏差),所以我基本上是随机创建新的观察结果与现有行的组合,这有望减少这种偏差。

init_train = structure(list(ID = structure(c(16L, 24L, 29L, 30L, 31L, 1L, 
2L, 3L, 4L, 5L, 6L, 7L, 8L, 9L, 10L, 11L, 12L, 13L, 14L, 15L, 
17L, 18L, 19L, 20L, 21L, 22L, 23L, 25L, 26L, 27L, 28L), .Label = c("SDataID10", 
"SDataID11", "SDataID13", "SDataID14", "SDataID15", "SDataID16", 
"SDataID17", "SDataID18", "SDataID19", "SDataID20", "SDataID21", 
"SDataID24", "SDataID25", "SDataID28", "SDataID29", "SDataID3", 
"SDataID31", "SDataID32", "SDataID34", "SDataID35", "SDataID37", 
"SDataID38", "SDataID39", "SDataID4", "SDataID43", "SDataID44", 
"SDataID45", "SDataID46", "SDataID5", "SDataID7", "SDataID8"), class = "factor"), 
    V1 = c(161L, 11L, 32L, 13L, 194L, 63L, 89L, 78L, 87L, 81L, 
    63L, 198L, 9L, 196L, 189L, 116L, 104L, 5L, 173L, 5L, 87L, 
    5L, 45L, 19L, 133L, 8L, 42L, 45L, 45L, 176L, 63L), V2 = structure(c(1L, 
    3L, 3L, 1L, 3L, 2L, 1L, 3L, 3L, 1L, 1L, 1L, 3L, 1L, 3L, 2L, 
    1L, 1L, 3L, 1L, 1L, 1L, 1L, 3L, 3L, 1L, 1L, 1L, 1L, 1L, 1L
    ), .Label = c("ONE", "THREE", "TWO"), class = "factor"), 
    V3 = c(1L, 2L, 2L, 1L, 2L, 3L, 1L, 2L, 2L, 1L, 3L, 3L, 3L, 
    2L, 2L, 3L, 1L, 2L, 3L, 3L, 3L, 2L, 1L, 2L, 2L, 1L, 1L, 1L, 
    1L, 1L, 1L), V5 = structure(c(1L, 3L, 1L, 3L, 1L, 1L, 1L, 
    1L, 3L, 3L, 1L, 3L, 3L, 3L, 2L, 4L, 1L, 2L, 1L, 2L, 1L, 3L, 
    1L, 3L, 1L, 3L, 3L, 3L, 1L, 1L, 3L), .Label = c("FOUR", "ONE", 
    "THREE", "TWO"), class = "factor"), V6 = c(0L, 2L, 2L, 2L, 
    0L, 0L, 0L, 0L, 1L, 0L, 0L, 0L, 0L, 2L, 1L, 0L, 0L, 3L, 0L, 
    3L, 3L, 1L, 0L, 0L, 0L, 0L, 0L, 1L, 0L, 0L, 3L), Target = c(0L, 
    1L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 
    0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L, 0L
    )), .Names = c("ID", "V1", "V2", "V3", "V5", "V6", "Target"
), class = "data.frame", row.names = c(NA, -31L))

set.seed(1000)
train = as.data.frame(lapply(init_train, function(x) sample(x, 27700, replace = TRUE)))