[英]How to define the classification threshold as a (hyper)parameter of a learner for tuning in mlr3 package in R?
there is a function to tune threshold for say a binary classification described here: https://mlr3pipelines.mlr-org.com/reference/mlr_pipeops_t.nethreshold.html有一个 function 来调整这里描述的二进制分类的阈值: https://mlr3pipelines.mlr-org.com/reference/mlr_pipeops_t.nethreshold.html
Here's my failed attempt:这是我失败的尝试:
RF_lrn <- lrn("classif.rfsrc", id = "rf", predict_type = "prob")
RF_lrn$param_set$values = list(na.action = "na.impute", seed = -123)
single_pred_rf = po("subsample", frac = 1, id = "rf_ss") %>>%
po("learner", RF_lrn) %>>% po("tunethreshold")
That did not work in my mlr3 pipeline and I did not find any solution explained anywhere so here is my code:这在我的 mlr3 管道中不起作用,我没有在任何地方找到任何解释的解决方案,所以这是我的代码:
xgb_lrn <-
lrn("classif.xgboost", id = "xgb", predict_type = "prob")
single_pred_xgb = po("subsample", frac = 1, id = "xgb_ss") %>>%
po("learner", xgb_lrn)
lrnrs <- list(
RF_lrn,
xgb_lrn)
lrnrs <- lapply(lrnrs, function(x) {
GraphLearner$new(po("learner_cv", x) %>>% po("tunethreshold",
param_vals = list(
measure = "classif.prauc"
)
))
})
library("GenSA")
lrnrs$RF_lrn <- auto_tuner(
learner = RF_lrn,
search_space = ps(
ntree = p_int(lower = 20, upper = 300),
mtry = p_int(lower = 2, upper = 5),
nodesize = p_int(lower = 1, upper = 7)
),
resampling = rsmp("bootstrap", repeats = 2, ratio = 0.8),
measure = msr("classif.prauc"),
term_evals = 100,
method = "random_search"
)
which somehow works but I want the decision threshold to be tuned as a parameter the same way I tune other hyperparameters using the random search in benchmarking/cross validation.它以某种方式起作用,但我希望将决策阈值作为参数进行调整,就像我在基准测试/交叉验证中使用随机搜索调整其他超参数一样。 Any solution?任何解决方案? Thanks in advance提前致谢
the solution is to use po("threshold")
instead of po("t.nethreshold")
as suggested in the comments and this mlr gallery example解决方案是使用po("threshold")
而不是评论中建议的po("t.nethreshold")
和这个mlr gallery example
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.