R語言實現隨機森林

轉自簡書
鏈接：http://www.jianshu.com/p/ca09dedb0541

1、讓兩個以及兩個以上組合樹變成一顆樹：combine()

combine(...)

…：每個隨機森林對象

data(iris)
rf1 <- randomForest(Species ~ ., iris, ntree=50, norm.votes=FALSE)
rf2 <- randomForest(Species ~ ., iris, ntree=50, norm.votes=FALSE)
rf3 <- randomForest(Species ~ ., iris, ntree=50, norm.votes=FALSE)
rf.all <- combine(rf1, rf2, rf3)
print(rf.all)

2、從森林中提取一顆樹：getTree()

getTree(rfobj, k=1, labelVar=FALSE)

rfobj：隨機森林對象
k：提取樹的個數
labelVar：FALSE or TRUE，更好的標籤被用於分裂變量和預測的類別
對於數值預測，數據與變量的值小於或等於分裂點去到左子節點。
對於分類的預測，分裂點代表一個整數，依據其二進制擴展可以判斷該類別是去左子節點還是去右子節點。
例如，如果一個預測四大類，分裂點爲13。13的二進制擴展是（1，0，1，1）（因爲
13 = 1(2^0)+0(2^1)+1(2^2)+1(2^3)），所以類別1、3或4被預測到左子節點，其餘的到右子節點。
一個矩陣（或數據框，如果labelvar = true）六列六行，總數等於樹中的節點數。六列是：

左子女：左子節點所在的行；如果節點爲終端的話，則爲0
右子女：右子節點所在的行；如果節點爲終端的話，則爲0
分裂的變量：被用來分割的節點；如果節點是終端，則爲0
分裂點：最好的分裂點
狀態：節點終端（- 1）或不（1）
預測：節點的預測；如果節點不是終端，則爲0

## Look at the third trees in the forest.
getTree(randomForest(iris[,-5], iris[,5], ntree=10), 3, labelVar=TRUE)

3、增加樹的數量：grow()

grow(x, how.many, ...)

x：隨機森林對象
how.many：增加樹的數量

data(iris)
iris.rf <- randomForest(Species ~ ., iris, ntree=50, norm.votes=FALSE)
iris.rf <- grow(iris.rf, 50)
print(iris.rf)

4、提取特徵的重要性：importance()

importance(x, type=NULL, class=NULL, scale=TRUE, ...)

x：隨機森林對象
type ：1或者2,重要性度量類型，1代表均值降序精度，2代表均值降序節點純度
class：類別度量

set.seed(4543)
data(mtcars)
mtcars.rf <- randomForest(mpg ~ ., data=mtcars, ntree=1000,
keep.forest=FALSE, importance=TRUE)
importance(mtcars.rf)
importance(mtcars.rf, type=1)

5、畫圖：plot()

plot(x, sort=TRUE, ...)
set.seed(1)
data(iris)
iris.rf <- randomForest(Species ~ ., iris, keep.forest=FALSE)
plot(margin(iris.rf))

6、分類圖：MDSplot()

MDSplot(rf, fac, k=2, palette=NULL, pch=20, ...)

rf：隨機森林對象
fac：訓練隨機森林的因子
k：維度數
palette：顏色

set.seed(1)
data(iris)
iris.rf <- randomForest(Species ~ ., iris, proximity=TRUE,
keep.forest=FALSE)
MDSplot(iris.rf, iris$Species)
## Using different symbols for the classes:
MDSplot(iris.rf, iris$Species, palette=rep(1, 3), pch=as.numeric(iris$Species))

7、填補缺失值的中位數：na.roughfix()

na.roughfix(object, ...)

數值變量：缺失值使用中位數替代
因子自變量：缺失值使用最常出現的替代

data(iris)
iris.na <- iris
set.seed(111)
iris.na
## artificially drop some data values.
for (i in 1:4) iris.na[sample(150, sample(20)), i] <- NA
iris.roughfix <- na.roughfix(iris.na)
iris.narf <- randomForest(Species ~ ., iris.na, na.action=na.roughfix)
print(iris.narf)

8、異常值（離羣點）：outlier()

outlier(x, ...)
set.seed(1)
iris.rf <- randomForest(iris[,-5], iris[,5], proximity=TRUE)
plot(outlier(iris.rf), type="h",
col=c("red", "green", "blue")[as.numeric(iris$Species)])

9、用圖形化描述變量的邊際效應（分類或迴歸）：partialPlot()

partialPlot(x, pred.data, x.var, which.class,
w, plot = TRUE, add = FALSE,
n.pt = min(length(unique(pred.data[, xname])), 51),
rug = TRUE, xlab=deparse(substitute(x.var)), ylab="",
main=paste("Partial Dependence on", deparse(substitute(x.var))),
...)

x：隨機森林對象
pred.data：預測數據
x.var：變量名稱
which.class：分類數據
w：權重

data(iris)
set.seed(543)
iris.rf <- randomForest(Species~., iris)
partialPlot(iris.rf, iris, Petal.Width, "versicolor")
## Looping over variables ranked by importance:
data(airquality)
airquality <- na.omit(airquality)
set.seed(131)
ozone.rf <- randomForest(Ozone ~ ., airquality, importance=TRUE)
imp <- importance(ozone.rf)
impvar <- rownames(imp)[order(imp[, 1], decreasing=TRUE)]
op <- par(mfrow=c(2, 3))
for (i in seq_along(impvar)) {
  partialPlot(ozone.rf, airquality, impvar[i], xlab=impvar[i],
              main=paste("Partial Dependence on", impvar[i]),
              ylim=c(30, 70),col='blue')
}
par(op)

10、畫均值方差：Plot()

plot(x, type="l", main=deparse(substitute(x)), ...)
data(mtcars)
plot(randomForest(mpg ~ ., mtcars, keep.forest=FALSE, ntree=100), log="y",col='red')

11、預測測試數據：predict()

predict(object, newdata, type="response",
norm.votes=TRUE, predict.all=FALSE, proximity=FALSE, nodes=FALSE,
cutoff, ...)

object：隨機森林對象
newdata：預測新的數據
type：使用概率矩陣或者還是使用計數投票矩陣
norm.votes：計數投票矩陣標準化
predict.all：是否使用所有的樹進行預測
nodes：最終端的節點
cutoff：

data(iris)
set.seed(111)
ind <- sample(2, nrow(iris), replace = TRUE, prob=c(0.8, 0.2))
iris.rf <- randomForest(Species ~ ., data=iris[ind == 1,])
iris.pred <- predict(iris.rf, iris[ind == 2,])
table(observed = iris[ind==2, "Species"], predicted = iris.pred)
## Get prediction for all trees.
predict(iris.rf, iris[ind == 2,], predict.all=TRUE)
## Proximities.
predict(iris.rf, iris[ind == 2,], proximity=TRUE)
## Nodes matrix.
str(attr(predict(iris.rf, iris[ind == 2,], nodes=TRUE), "nodes"))

12、隨機森林模型：randomForest()

## S3 method for class ’formula’
randomForest(formula, data=NULL, ..., subset, na.action=na.fail)
## Default S3 method:
randomForest(x, y=NULL, xtest=NULL, ytest=NULL, ntree=500,
mtry=if (!is.null(y) && !is.factor(y))
max(floor(ncol(x)/3), 1) else floor(sqrt(ncol(x))),
replace=TRUE, classwt=NULL, cutoff, strata,
sampsize = if (replace) nrow(x) else ceiling(.632*nrow(x)),
nodesize = if (!is.null(y) && !is.factor(y)) 5 else 1,
maxnodes = NULL,
importance=FALSE, localImp=FALSE, nPerm=1,
proximity, oob.prox=proximity,
norm.votes=TRUE, do.trace=FALSE,
keep.forest=!is.null(y) && is.null(xtest), corr.bias=FALSE,
keep.inbag=FALSE, ...)

## S3 method for class ’randomForest’
print(x, ...)

分類:

> data(iris)
> set.seed(71)
> iris.rf <- randomForest(Species ~ ., data=iris, importance=TRUE,
+ proximity=TRUE)
> print(iris.rf)
Call:
 randomForest(formula = Species ~ ., data = iris, importance = TRUE,      proximity = TRUE) 
               Type of random forest: classification
                     Number of trees: 500
No. of variables tried at each split: 2

        OOB estimate of  error rate: 5.33%
Confusion matrix:
           setosa versicolor virginica class.error
setosa         50          0         0        0.00
versicolor      0         46         4        0.08
virginica       0          4        46        0.08
> round(importance(iris.rf), 2)
             setosa versicolor virginica MeanDecreaseAccuracy
Sepal.Length   6.04       7.85      7.93                11.51
Sepal.Width    4.40       1.03      5.44                 5.40
Petal.Length  21.76      31.33     29.64                32.94
Petal.Width   22.84      32.67     31.68                34.50
             MeanDecreaseGini
Sepal.Length             8.77
Sepal.Width              2.19
Petal.Length            42.54
Petal.Width             45.77
> iris.mds <- cmdscale(1 - iris.rf$proximity, eig=TRUE)
> op <- par(pty="s")
> pairs(cbind(iris[,1:4], iris.mds$points), cex=0.6, gap=0,
+ col=c("red", "green", "blue")[as.numeric(iris$Species)],
+ main="Iris Data: Predictors and MDS of Proximity Based on RandomForest")
> par(op)

無監督案例:

> set.seed(17)
> iris.urf <- randomForest(iris[, -5])
> MDSplot(iris.urf, iris$Species)
> ## stratified sampling: draw 20, 30, and 20 of the species to grow each tree.
> (iris.rf2 <- randomForest(iris[1:4], iris$Species,
+ sampsize=c(20, 30, 20)))

Call:
 randomForest(x = iris[1:4], y = iris$Species, sampsize = c(20,      30, 20)) 
               Type of random forest: classification
                     Number of trees: 500
No. of variables tried at each split: 2

        OOB estimate of  error rate: 5.33%
Confusion matrix:
           setosa versicolor virginica class.error
setosa         50          0         0        0.00
versicolor      0         47         3        0.06
virginica       0          5        45        0.10

迴歸:

data(airquality)
set.seed(131)
ozone.rf <- randomForest(Ozone ~ ., data=airquality, mtry=3,
                         importance=TRUE, na.action=na.omit)
print(ozone.rf)
## Show "importance" of variables: higher value mean more important:
round(importance(ozone.rf), 2)
## "x" can be a matrix instead of a data frame:
set.seed(17)
x <- matrix(runif(5e2), 100)
y <- gl(2, 50)
(myrf <- randomForest(x, y))
(predict(myrf, x))
## "complicated" formula:
(swiss.rf <- randomForest(sqrt(Fertility) ~ . - Catholic + I(Catholic < 50),
                          data=swiss))
(predict(swiss.rf, swiss))
## Test use of 32-level factor as a predictor:
set.seed(1)
x <- data.frame(x1=gl(53, 10), x2=runif(530), y=rnorm(530))
(rf1 <- randomForest(x[-3], x[[3]], ntree=10))
## Grow no more than 4 nodes per tree:
(treesize(randomForest(Species ~ ., data=iris, maxnodes=4, ntree=30)))
## test proximity in regression
iris.rrf <- randomForest(iris[-1], iris[[1]], ntree=101, proximity=TRUE, oob.prox=FALSE)
str(iris.rrf$proximity)

13、交叉驗證進行特徵選擇：rfcv()

rfcv(trainx, trainy, cv.fold=5, scale="log", step=0.5,
mtry=function(p) max(1, floor(sqrt(p))), recursive=FALSE, ...)
trainx：自變量
trainy：因變量
cv.fold：幾折交叉驗證
set.seed(647)
myiris <- cbind(iris[1:4], matrix(runif(96 * nrow(iris)), nrow(iris), 96))
result <- rfcv(myiris, iris$Species, cv.fold=3)
with(result, plot(n.var, error.cv, log="x", type="o", lwd=2))
## The following can take a while to run, so if you really want to try
## it, copy and paste the code into R.
## Not run:
result <- replicate(5, rfcv(myiris, iris$Species), simplify=FALSE)
error.cv <- sapply(result, "[[", "error.cv")
matplot(result[[1]]$n.var, cbind(rowMeans(error.cv), error.cv), type="l",
lwd=c(2, rep(1, ncol(error.cv))), col=1, lty=1, log="x",
xlab="Number of variables", ylab="CV Error")
## End(Not run)

14、隨機森林填補缺失值：rfImpute ()

## Default S3 method:
rfImpute(x, y, iter=5, ntree=300, ...)

## S3 method for class ’formula’
rfImpute(x, data, ..., subset)

x：自變量
y：因變量

data：數據框
data(iris)
iris.na <- iris
set.seed(111)
## artificially drop some data values.
for (i in 1:4) iris.na[sample(150, sample(20)), i] <- NA
set.seed(222)
iris.imputed <- rfImpute(Species ~ ., iris.na)
set.seed(333)
iris.rf <- randomForest(Species ~ ., iris.imputed)
print(iris.rf)

15、最優抽樣：rfImpute()

tuneRF(x, y, mtryStart, ntreeTry=50, stepFactor=2, improve=0.05,
trace=TRUE, plot=TRUE, doBest=FALSE, ...)
data(fgl, package="MASS")
fgl.res <- tuneRF(fgl[,-10], fgl[,10], stepFactor=1.5)

16、變量的重要性：varImpPlot()

arImpPlot(x, sort=TRUE, n.var=min(30, nrow(x$importance)),
type=NULL, class=NULL, scale=TRUE,
main=deparse(substitute(x)), ...)

set.seed(4543)
data(mtcars)
mtcars.rf <- randomForest(mpg ~ ., data=mtcars, ntree=1000, keep.forest=FALSE,
                          importance=TRUE)
varImpPlot(mtcars.rf)

隨機森林中決策樹分裂方法

https://www.stat.berkeley.edu/~breiman/RandomForests/cc_home.htm
Gini importance
Every time a split of a node is made on variable m the gini impurity criterion for the two descendent nodes is less than the parent node. Adding up the gini decreases for each individual variable over all trees in the forest gives a fast variable importance that is often very consistent with the permutation importance measure.

變量重要性評價

隨機森林變量重要性的計算方法有兩種，分別是Gini指數和測試集(OOB)錯誤率。

importance(mtcars.rf)
       %IncMSE IncNodePurity
cyl  16.799437     173.03496
disp 18.946107     241.43741
hp   17.282802     186.55081
drat  6.961155      70.14317
wt   19.012343     248.53222
qsec  5.179746      30.64678
vs    5.147341      31.76982
am    5.357654      18.35507
gear  4.324805      16.00897
carb  9.825615      27.77433

其中，%IncMSE和IncNodePurity分別是相對重要性和節點純度，其中，相對重要性基於OOB錯誤率，節點純度基於Gini指數。