I have data that has an ID column in it. I drop this column from my trainTask as it is not a feature. However, I would like to link the prediction probability up with the actual ID number in the data.
The column I want to match on is Init_Acct which is the ID number in the data.frame
My code is below:
# Make classif tasks
trainTask <- makeClassifTask(
data = train.df %>% dplyr::select(-Init_Acct) # Init_Acct is the ID I want to match
, id
, target = "READMIT_FLAG"
, positive = "Y"
)
testTask <- makeClassifTask(
data = test.df %>% dplyr::select(-Init_Acct)
, target = "READMIT_FLAG"
, positive = "Y"
)
# Check trainTask and testTask
trainTask <- smote(trainTask, rate = 6)
testTask <- smote(testTask, rate = 6)
# GBM ####
getParamSet('classif.gbm')
gbm.learner <- makeLearner(
'classif.gbm'
, predict.type = 'prob'
)
plotLearnerPrediction(gbm.learner, trainTask)
# Tune model
gbm.tune.ctl <- makeTuneControlRandom(maxit = 50L)
# Cross validation
gbm.cv <- makeResampleDesc("CV", iters = 3L)
# Grid search - Hyper-parameter space
gbm.par <- makeParamSet(
makeDiscreteParam('distribution', values = 'bernoulli')
, makeIntegerParam('n.trees', lower = 10, upper = 1000)
, makeIntegerParam('interaction.depth', lower = 2, upper = 10)
, makeIntegerParam('n.minobsinnode', lower = 10, upper = 80)
, makeNumericParam('shrinkage', lower = 0.01, upper = 1)
)
# Tune Hyper-parameters
parallelMap::parallelStartSocket(
4
, level = "mlr.tuneParams"
)
gbm.tune <- tuneParams(
learner = gbm.learner
, task = trainTask
, resampling = gbm.cv
, measures = acc
, par.set = gbm.par
, control = gbm.tune.ctl
)
parallelMap::parallelStop()
# Check CV acc
gbm.tune$y
gbm.tune$x
# Set hyper-parameters
gbm.ps <- setHyperPars(
learner = gbm.learner
, par.vals = gbm.tune$x
)
# Train gbm
gbm.train <- train(gbm.ps, testTask)
plotLearningCurve(
generateLearningCurveData(
gbm.learner
, testTask
)
)
# Predict
gbm.pred <- predict(gbm.train, testTask)
plotResiduals(gbm.pred)
# Create submission file
gbm.submit <- data.frame(
gbm.pred$data
)
head(gbm.submit, 5)
table(gbm.submit$truth, gbm.submit$response)
# Confusion Matrix
calculateConfusionMatrix(gbm.pred)
calculateROCMeasures(gbm.pred)
conf_mat_f1_func(gbm.pred)
perf_plots_func(Model = gbm.pred)
Data would look something like this:
glimpse(train.df)
Observations: 33,031
Variables: 17
$ Init_Acct <chr> "12345678", "87654321", "81734650", "11223344", "1422...
$ Init_LOS <dbl> 2, 2, 5, 1, 12, 3, 16, 9, 3, 14, 1, 1, 4, 7, 4, 1, 3,...
$ Init_LACE <dbl> 2, 7, 7, 9, 8, 8, 11, 10, 8, 10, 5, 4, 8, 8, 4, 5, 3,...
$ READMIT_FLAG <fct> N, N, N, N, N, N, N, N, N, N, N, N, N, N, N, N, N, Y,...
$ Init_Hosp_Pvt <fct> PRIVATE, HOSPITALIST, HOSPITALIST, HOSPITALIST, PRIVA...
$ Age_at_Init_Admit <dbl> 37, 26, 56, 67, 51, 53, 48, 57, 92, 67, 72, 22, 60, 6...
$ Age_Bucket <fct> 3, 2, 5, 6, 5, 5, 4, 5, 9, 6, 7, 2, 6, 6, 7, 6, 9, 5,...
$ Gender <fct> F, M, F, M, M, F, M, F, M, M, M, F, M, F, F, F, F, M,...
$ Init_ROM <dbl> 1, 1, 3, 4, 2, 3, 1, 3, 4, 1, 1, 1, 1, 1, 2, 1, 2, 4,...
$ Init_SOI <dbl> 1, 1, 3, 4, 3, 3, 3, 3, 3, 2, 1, 1, 2, 2, 2, 2, 2, 4,...
$ Has_Diabetes <fct> N, N, N, N, N, N, N, N, N, N, N, N, N, N, N, N, N, N,...
$ reduced_dispo <fct> AHR, AHR, AHR, ATH, ATW, ATW, ATW, AHR, AHR, ATW, AHR...
$ reduced_hsvc <fct> SUR, MED, MED, Other, MED, MED, MED, MED, MED, MED, M...
$ reduced_abucket <fct> 3, 2, 5, 6, 5, 5, 4, 5, Other, 6, 7, 2, 6, 6, 7, 6, O...
$ reduced_spclty <fct> Other, HOSIM, HOSIM, HOSIM, Other, HOSIM, HOSIM, HOSI...
$ reduced_lihn <fct> Other, Medical, Pneumonia, Medical, Medical, Medical,...
$ discharge_month <fct> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,...
The output:
glimpse(gbm.submit)
Observations: 23,896
Variables: 5
$ id <int> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18,...
$ truth <fct> Y, N, N, N, N, N, N, Y, N, N, N, N, N, N, Y, N, Y, N, N, N, N,...
$ prob.N <dbl> 0.9150623, 0.7914781, 0.9661108, 0.9198683, 0.8502536, 0.94376...
$ prob.Y <dbl> 0.08493774, 0.20852192, 0.03388919, 0.08013167, 0.14974644, 0....
$ response <fct> N, N, N, N, N, N, N, Y, N, N, N, N, N, N, N, N, Y, N, N, N, N,...
MLR's predict() preserves row names and produces an additional id column in its output that indexes the original data. You can use either one to associate predictions with their original sample IDs.
Setup
library(tidyverse)
library(mlr)
## Add a custom sample ID column
iris2 <- iris %>% mutate(Init_Acct = paste0("Acct",1:n()))
lrn <- makeLearner( "classif.gbm", predict.type="prob" )
Option 1: Using id column to index the original data
## Drop the custom column as in your original post
task <- makeClassifTask( data=select(iris2, -Init_Acct), target="Species" )
mdl <- train( lrn, task )
pred <- predict( mdl, task )
## Join against the original data by the "id" column
iris2 %>% mutate(id=1:n()) %>% select(Init_Acct, id) %>%
inner_join( pred$data ) %>% select(-id)
# Init_Acct truth prob.setosa prob.versicolor prob.virginica response
# 1 Acct1 setosa 0.9998775 1.225043e-04 2.836942e-08 setosa
# 2 Acct2 setosa 0.9999652 3.468690e-05 1.118015e-07 setosa
# 3 Acct3 setosa 0.9999538 4.611200e-05 8.389636e-08 setosa
Option 2: Use rownames
## Store the sample names into rownames
task <- makeClassifTask( data=column_to_rownames(iris2, "Init_Acct"),
target="Species" )
mdl <- train( lrn, task )
pred <- predict( mdl, task )
## Pull the rownames back out into their own column
pred$data %>% rownames_to_column( "Init_Acct" ) %>% select(-id)
# Init_Acct truth prob.setosa prob.versicolor prob.virginica response
# 1 Acct1 setosa 9.999266e-01 7.331226e-05 6.889259e-08 setosa
# 2 Acct2 setosa 9.999751e-01 2.462816e-05 3.154618e-07 setosa
# 3 Acct3 setosa 9.999656e-01 3.421543e-05 1.449155e-07 setosa