feature request: support for classprobs with bstSm
crossxwill opened this issue · 0 comments
crossxwill commented
The method bstSm
does not appear to support classProbs=TRUE. I'm requesting that the method support probability predictions. Here's a reprex that shows the error when class probabilities are requested for bstSm
.
library(caret)
## data
set.seed(1)
x <- rnorm(1000)
z <- -2 + 2*x + rnorm(1000)
y <- rbinom(1000, 1, boot::inv.logit(z))
df <- data.frame(y=y, x=x)
summary(df)
## response
df$label <- factor(ifelse(df$y == 1, "yes", "no"),
levels=c("yes","no"))
summary(df)
## control
fcstHorizon <- 100
initWindow <- 800
param_skip <- fcstHorizon - 1
fitControl_oneSE <- trainControl(method = "timeslice",
initialWindow=initWindow,
horizon=fcstHorizon,
fixedWindow=FALSE,
skip=param_skip,
## Estimate class probabilities
classProbs = TRUE,
## Evaluate performance using
## the following function
summaryFunction = mnLogLoss,
selectionFunction="oneSE")
## gamboost
set.seed(1)
gam_mod <- train(
label ~ x,
data = df,
method = "gamboost",
trControl = fitControl_oneSE,
metric = "logLoss",
dfbase =3
)
plot(gam_mod)
## bstSm (errors out)
set.seed(1)
bst_mod <- train(
label ~ x,
data = df,
method = "bstSm",
trControl = fitControl_oneSE,
metric = "logLoss"
)
plot(bst_mod)