机器算法验证 - RandomForest 和类权重 - 吾爱随笔录

一句话问题：有人知道如何为随机森林确定好的类权重吗？

说明：我在玩不平衡的数据集。我想使用这个R包randomForest来在一个非常倾斜的数据集上训练一个模型，只有很少的正面例子和很多负面的例子。我知道，还有其他方法，最后我会使用它们，但出于技术原因，构建随机森林是一个中间步骤。所以我玩弄了参数classwt。我在半径为 2 的圆盘中设置了一个非常人工的 5000 个负样本数据集，然后我在半径为 1 的圆盘中采样了 100 个正样本。我怀疑的是

1) 没有类权重，模型会变得“退化”，即FALSE到处都可以预测。

2）使用公平的类权重，我会在中间看到一个“绿点”，即它会预测半径为 1 的圆盘，TRUE尽管有负样本。

这是数据的样子：

这是在没有加权的情况下发生的情况：（调用是randomForest(x = train[, .(x,y)],y = as.factor(train$z),ntree = 50)：）

为了检查，我还尝试了当我通过对负类进行下采样以使关系再次为 1:1 来猛烈地平衡数据集时会发生什么。这给了我预期的结果：

但是，当我计算一个类别权重为“FALSE”= 1、“TRUE”= 50 的模型时（这是一个公平的权重，因为负数比正数多 50 倍），然后我得到：

只有当我将权重设置为像 'FALSE' = 0.05 和 'TRUE' = 500000 这样的奇怪值时，我才会得到有意义的结果：

这是非常不稳定的，即将“FALSE”权重更改为 0.01 会使模型再次退化（即它TRUE在任何地方都可以预测）。

问题：有人知道如何为随机森林确定好的类权重吗？

代码：

library(plot3D)
library(data.table)
library(randomForest)
set.seed(1234)
amountPos = 100
amountNeg = 5000

# positives
r = runif(amountPos, 0, 1)
phi = runif(amountPos, 0, 2*pi)
x = r*cos(phi)
y = r*sin(phi)
z = rep(T, length(x))
pos = data.table(x = x, y = y, z = z)

# negatives
r = runif(amountNeg, 0, 2)
phi = runif(amountNeg, 0, 2*pi)
x = r*cos(phi)
y = r*sin(phi)
z = rep(F, length(x))
neg = data.table(x = x, y = y, z = z)

train = rbind(pos, neg)

# draw train set, verify that everything looks ok
plot(train[z == F]$x, train[z == F]$y, col="red")
points(train[z == T]$x, train[z == T]$y, col="green")
# looks ok to me :-)

Color.interpolateColor = function(fromColor, toColor, amountColors = 50) {
  from_rgb = col2rgb(fromColor)
  to_rgb = col2rgb(toColor)

  from_r = from_rgb[1,1]
  from_g = from_rgb[2,1]
  from_b = from_rgb[3,1]

  to_r = to_rgb[1,1]
  to_g = to_rgb[2,1]
  to_b = to_rgb[3,1]

  r = seq(from_r, to_r, length.out = amountColors)
  g = seq(from_g, to_g, length.out = amountColors)
  b = seq(from_b, to_b, length.out = amountColors)

  return(rgb(r, g, b, maxColorValue = 255))
}
DataTable.crossJoin = function(X,Y) {
  stopifnot(is.data.table(X),is.data.table(Y))
  k = NULL
  X = X[, c(k=1, .SD)]
  setkey(X, k)
  Y = Y[, c(k=1, .SD)]
  setkey(Y, k)
  res = Y[X, allow.cartesian=TRUE][, k := NULL]
  X = X[, k := NULL]
  Y = Y[, k := NULL]
  return(res)
}

drawPredictionAreaSimple = function(model) {
  widthOfSquares = 0.1
  from = -2
  to = 2

  xTable = data.table(x = seq(from=from+widthOfSquares/2,to=to-widthOfSquares/2,by = widthOfSquares))
  yTable = data.table(y = seq(from=from+widthOfSquares/2,to=to-widthOfSquares/2,by = widthOfSquares))
  predictionTable = DataTable.crossJoin(xTable, yTable)
  pred = predict(model, predictionTable)
  res = rep(NA, length(pred))
  res[pred == "FALSE"] = 0
  res[pred == "TRUE"] = 1
  pred = res
  predictionTable = predictionTable[, PREDICTION := pred]
  #predictionTable = predictionTable[y == -1 & x == -1, PREDICTION := 0.99]
  col = Color.interpolateColor("red", "green")

  input = matrix(c(predictionTable$x, predictionTable$y), nrow = 2, byrow = T)
  m = daply(predictionTable, .(x, y), function(x) x$PREDICTION)
  image2D(z = m, x = sort(unique(predictionTable$x)), y = sort(unique(predictionTable$y)), col = col, zlim = c(0,1))
}


rfModel = randomForest(x = train[, .(x,y)],y = as.factor(train$z),ntree = 50)
rfModelBalanced = randomForest(x = train[, .(x,y)],y = as.factor(train$z),ntree = 50, classwt = c("FALSE" = 1, "TRUE" = 50))
rfModelBalancedWeird = randomForest(x = train[, .(x,y)],y = as.factor(train$z),ntree = 50, classwt = c("FALSE" = 0.05, "TRUE" = 500000))


drawPredictionAreaSimple(rfModel)
title("unbalanced")
drawPredictionAreaSimple(rfModelBalanced)
title("balanced with weights")
pos = train[z == T]
neg = train[z == F]
neg = neg[sample.int(neg[, .N], size = 100, replace = FALSE)]
trainSampled = rbind(pos, neg)
rfModelBalancedSampling = randomForest(x = trainSampled[, .(x,y)],y = as.factor(trainSampled$z),ntree = 50)
drawPredictionAreaSimple(rfModelBalancedSampling)
title("balanced with sampling")


drawPredictionAreaSimple(rfModelBalancedWeird)
title("balanced with weird weights")