我有一个非常不平衡的数据集。为了解决这个问题,我分别尝试了不同的类不平衡技术:downSample、类权重、阈值调整。其中,阈值调整效果最差。单独使用 downSample 或单独使用类权重,我没有设法获得足够好的结果:要么有太多的 FalsePositives 要么 FalseNegatives。所以我想将这两种技术结合起来。这是我累的:
# produce some re-producible imbalanced data
set.seed(12345)
y <- as.factor(sample(c("M", "F"),
prob = c(0.1, 0.9),
size = 10000,
replace = TRUE))
x <- rnorm(10000)
DATA <- data.frame(y = as.factor(y), x)
set.seed(12345)
folds <- createFolds(dataSet$y, k = 10,
list = TRUE, returnTrain = TRUE)
# class weights
k <- 0.5
classWeights <- ifelse(DATA$y == "M",
(1/table(DATA$y)[1]) * k,
(1/table(DATA$y)[2]) * (1-k))
所以当我不把sampling
论点放在controlTrain
:
# select algorithm
algorithm <- "bayesglm"
# train parameters
set.seed(12345)
traincontrol <- trainControl(method = "loocv", # resampling method
number = 10,
index = folds,
classProbs = TRUE,
summaryFunction = twoClassSummary,
savePredictions = TRUE,
# sampling = "down"
)
fitModel <- train(y ~ .,
data = DATA,
trControl = traincontrol,
method = algorithm,
metric = "ROC",
weights = classWeights,
)
它有效并且没有错误。但是当我将 trainControl 的采样参数添加为
# train parameters
set.seed(12345)
traincontrol <- trainControl(method = "loocv", # resampling method
number = 10,
index = folds,
classProbs = TRUE,
summaryFunction = twoClassSummary,
savePredictions = TRUE,
sampling = "down"
)
fitModel <- train(y ~ .,
data = DATA,
trControl = traincontrol,
method = algorithm,
metric = "ROC",
weights = classWeights,
)
我得到这个可以理解的错误:
Error in model.frame.default(formula = .outcome ~ ., data = list(x = c(-0.0640913631047556, :
variable lengths differ (found for '(weights)')
In addition: There were 11 warnings (use warnings() to see them)
Timing stopped at: 0.112 0.001 0.115
有没有办法做到这一点caret
?提前谢谢了。