Last updated: 2021-11-24
Checks: 7 0
Knit directory: spanish_data/
This reproducible R Markdown analysis was created with workflowr (version 1.6.2). The Checks tab describes the reproducibility checks that were applied when the results were created. The Past versions tab lists the development history.
Great! Since the R Markdown file has been committed to the Git repository, you know the exact version of the code that produced these results.
Great job! The global environment was empty. Objects defined in the global environment can affect the analysis in your R Markdown file in unknown ways. For reproduciblity it’s best to always run the code in an empty environment.
The command set.seed(20210329) was run prior to running the code in the R Markdown file. Setting a seed ensures that any results that rely on randomness, e.g. subsampling or permutations, are reproducible.
Great job! Recording the operating system, R version, and package versions is critical for reproducibility.
Nice! There were no cached chunks for this analysis, so you can be confident that you successfully produced the results during this run.
Great job! Using relative paths to the files within your workflowr project makes it easier to run your code on other machines.
Great! You are using Git for version control. Tracking code development and connecting the code version to the results is critical for reproducibility.
The results in this page were generated with repository version 6d4201c. See the Past versions tab to see a history of the changes made to the R Markdown and HTML files.
Note that you need to be careful to ensure that all relevant files for the analysis have been committed to Git prior to generating the results (you can use wflow_publish or wflow_git_commit). workflowr only checks the R Markdown file, but you know if there are other scripts or data files that it depends on. Below is the status of the Git repository when the results were generated:
Ignored files:
Ignored: .Rproj.user/
Ignored: manuscript/
Ignored: output/ap_models.RDS
Ignored: output/ap_result.RDS
Ignored: output/dis_models.RDS
Ignored: output/dis_result.RDS
Ignored: output/np_models.RDS
Ignored: output/np_result.RDS
Ignored: output/report_files/
Ignored: output/resample_perf.RDS
Ignored: output/resampling_models.RDS
Ignored: output/stepwise_performance.RDS
Note that any generated files, e.g. HTML, png, CSS, etc., are not included in this status report because it is ok for generated content to have uncommitted changes.
These are the previous versions of the repository in which changes were made to the R Markdown (analysis/6-codesPublish.Rmd) and HTML (docs/6-codesPublish.html) files. If you’ve configured a remote Git repository (see ?wflow_git_remote), click on the hyperlinks in the table below to view the files as they were in that past version.
| File | Version | Author | Date | Message |
|---|---|---|---|---|
| Rmd | be89c8d | Bernard | 2021-11-24 | new codes |
# Import packages --------------------------------------------------------------
# Helper
library(tidyverse)
library(data.table)
library (cowplot)
library (officer)
library (flextable)
# ML
library(mlr3)
library(mlr3learners)
library(mlr3tuning)
library(mlr3viz)
library(mlr3fselect)
library(mlr3pipelines)
library(mlr3hyperband)
library(iml)
library(MASS)
library(MLmetrics)
# Parallel
library (future)# Import data ------------------------------------------------------------------
dat <- readRDS("output/df.RDS")
np_dat <- bind_rows(dat$df_list$np$train_imp, dat$df_list$np$test_imp) #%>% group_by(outcome) %>% sample_n(50)
ap_dat <- bind_rows(dat$df_list$ap$train_imp, dat$df_list$ap$test_imp) #%>% group_by(outcome) %>% sample_n(50)
dis_dat <- bind_rows(dat$df_list$dis$train_imp, dat$df_list$dis$test_imp) #%>% group_by(outcome) %>% sample_n(50)
# Set ML task ------------------------------------------------------------------
task_np <- TaskClassif$new (id = "neckpain", backend = np_dat , target = "outcome")
task_np$nrow
task_np$feature_names
task_np$set_col_roles("ID", roles = "name")
task_ap <- TaskClassif$new (id = "armpain", backend = ap_dat , target = "outcome")
task_ap$nrow
task_ap$feature_names
task_ap$set_col_roles("ID", roles = "name")
task_dis <- TaskClassif$new (id = "disability", backend = dis_dat , target = "outcome")
task_dis$nrow
task_dis$feature_names
task_dis$set_col_roles("ID", roles = "name")
# Set preprocessing pipes-------------------------------------------------------
poe <- po("encode", method = "one-hot")
poscale <- po("scale", param_vals = list (center = TRUE, scale = TRUE))
# Set generic params------------------------------------------------------------
create_at = function(pipe_lrn, param_set){
auto_tuner(
method = "random_search",
learner = GraphLearner$new(pipe_lrn, task_type = "classif"),
resampling = rsmp("cv", folds = 3),
measure = msr("classif.auc"),
search_space = param_set,
term_evals = 2000
)
}# Define ML models -------------------------------------------------------------
######################### Logistic regression ##################################
lrn_logreg <- lrn("classif.log_reg", id = "log", predict_type = "prob")
grln_logreg <- poscale %>>%
lrn_logreg
logreg_gl = GraphLearner$new(grln_logreg, task_type = "classif")
######################### KNN ##################################################
lrn_kknn <- lrn("classif.kknn", id = "kknn", predict_type = "prob")
grln_kknn <- poscale %>>%
poe %>>%
lrn_kknn
ps_knn <- ParamSet$new(
params = list(
ParamInt$new("kknn.k", lower = 1, upper = 10))
)
kknn_at = create_at(grln_kknn, ps_knn)
######################### Xgboost ##############################################
lrn_xgb <- lrn("classif.xgboost", id = "xgb", predict_type = "prob", eta = 0.01)
grln_xgb <- poscale %>>%
poe %>>%
lrn_xgb
ps_xgb = ParamSet$new(
params = list(
ParamDbl$new("xgb.eta", lower = 0.001, upper = 0.2),
ParamDbl$new("xgb.max_depth", lower = 1, upper = 10),
ParamDbl$new("xgb.nrounds", lower = 50, upper = 500, tags = "budget"),
ParamDbl$new("xgb.colsample_bytree", lower = 0.5, upper = 1),
ParamDbl$new("xgb.colsample_bylevel", lower = 0.5, upper = 1),
ParamDbl$new("xgb.subsample", lower = 0.5, upper = 1),
ParamDbl$new("xgb.gamma", lower = -7, upper = 6),
ParamDbl$new("xgb.lambda", lower = -10, upper = 10),
ParamDbl$new("xgb.alpha", lower = -10, upper = 10)
))
ps_xgb$trafo = function(x, param_set) {
idx_gamma = grep("gamma", names(x))
x[[idx_gamma]] = 2^(x[[idx_gamma]])
idx_lambda = grep("lambda", names(x))
x[[idx_lambda]] = as.integer (2^(x[[idx_lambda]]))
idx_alpha = grep("alpha", names(x))
x[[idx_alpha]] = as.integer (2^(x[[idx_alpha]]))
idx_nrounds = grep("nrounds", names(x))
x[[idx_nrounds]] = as.integer (x[[idx_nrounds]])
idx_depth = grep("depth", names(x))
x[[idx_depth]] = as.integer (x[[idx_depth]])
x
}
xgb_at = create_at(grln_xgb, ps_xgb)
######################### Lasso ################################################
lrn_lasso <- lrn("classif.glmnet", id = "lasso", predict_type = "prob")
grln_lasso <- poscale %>>%
poe %>>%
lrn_lasso
ps_lasso = ParamSet$new(
params = list(ParamDbl$new("lasso.s", lower = 0, upper = 1))
)
lasso_at = create_at(grln_lasso, ps_lasso)
######################### Random forest ########################################
lrn_rf <- lrn("classif.ranger", id = "rf", predict_type = "prob")
grln_rf <- poscale %>>%
poe %>>%
lrn_rf
ps_rf <- ParamSet$new(list(
ParamInt$new ("rf.mtry", lower = 5, upper = 15, tags = "budget"),
ParamDbl$new ("rf.sample.fraction", lower = 0.5, upper = 1),
ParamInt$new ("rf.min.node.size", lower = 1, upper = 20)
))
rf_at = create_at(grln_rf, ps_rf)
######################### Nnet #################################################
lrn_net <- lrn("classif.nnet", id = "nnet", predict_type = "prob")
grln_net <- poscale %>>%
poe %>>%
lrn_net
ps_nnet <- ParamSet$new(list (
ParamInt$new ("nnet.size", lower = 1, upper = 10),
ParamDbl$new ("nnet.decay", lower = 0.1, upper = 0.5)
))
nnet_at = create_at(grln_net, ps_nnet)
######################### SVM ##################################################
lrn_svm <- lrn("classif.svm", id = "svm", type = "C-classification",
kernel = "radial", predict_type = "prob")
grln_svm <- poscale %>>%
poe %>>%
lrn_svm
ps_svm <- ParamSet$new(list (
ParamDbl$new ("svm.cost", lower = 0.1, upper = 10),
ParamDbl$new ("svm.gamma", lower = 0, upper = 5)
))
svm_at = create_at(grln_svm, ps_svm)
# Compare against stepwise model -----------------------------------------------
split_fit_predict <- function(data_index)
{
# recover splits
index <- bmr$resamplings$resampling[[data_index]]$instance
# define data
dat <- list(np_dat, ap_dat, dis_dat)[[data_index]]
dat_train <- dat[index$train,]
dat_test <- dat[index$test,]
# train models
mod_base <- glm(outcome ~ 1 + ., family = binomial(), data = dat_train[,-1])
mod_step <- stepAIC(mod_base)
best_model <- mod_step$formula
best_fitted <- mod_step$fitted.values
best_pred <- predict(mod_step, dat_test, type = "response")
auc <- MLmetrics::AUC (factor((best_fitted > 0.5) * 1), dat_train$outcome)
acc <- MLmetrics::Accuracy (factor((best_fitted > 0.5) * 1), dat_train$outcome)
sens <- MLmetrics::Sensitivity (dat_train$outcome, factor((best_fitted > 0.5) * 1))
spec <- MLmetrics::Specificity (dat_train$outcome, factor((best_fitted > 0.5) * 1))
prec <- MLmetrics::Precision (dat_train$outcome, factor((best_fitted > 0.5) * 1))
fbeta <- MLmetrics::FBeta_Score (dat_train$outcome, factor((best_fitted > 0.5) * 1))
metrics_train <- data.frame (
auc = auc,
acc = acc,
sens = sens,
spec = spec,
fbeta = fbeta
)
auc <- MLmetrics::AUC (factor((best_pred > 0.5) * 1), dat_test$outcome)
acc <- MLmetrics::Accuracy (factor((best_pred > 0.5) * 1), dat_test$outcome)
sens <- MLmetrics::Sensitivity (dat_test$outcome, factor((best_pred > 0.5) * 1))
spec <- MLmetrics::Specificity (dat_test$outcome, factor((best_pred > 0.5) * 1))
prec <- MLmetrics::Precision (dat_test$outcome, factor((best_pred > 0.5) * 1))
fbeta <- MLmetrics::FBeta_Score (dat_test$outcome, factor((best_pred > 0.5) * 1))
metrics_test <- data.frame (
auc = auc,
acc = acc,
sens = sens,
spec = spec,
prec = prec,
fbeta = fbeta
)
return(list(best_model = best_model,
metrics_train = metrics_train,
metrics_test = metrics_test))
}
res_step <- lapply(1:3, split_fit_predict)
names(res_step) <- c("np", "ap", "dis")
# extract best models
lapply(res_step, "[[", "best_model")
# extract metrics
lapply(res_step, "[[", "metrics_train")
lapply(res_step, "[[", "metrics_test")# Resampling performance -------------------------------------------------------
# See ## See https://mlr3gallery.mlr-org.com/posts/2020-07-27-bikesharing-demand/
## Method 1 - quick benchmarking
### Create benchmark grid
# Ability to do all tasks at once
#
tasks <- list (task_np, task_ap, task_dis)
learners <- list (logreg_gl, kknn_at, xgb_at, lasso_at, rf_at, nnet_at, svm_at)
design <- benchmark_grid(tasks = tasks,
learners = learners,
resamplings = rsmp("holdout", ratio = 0.8))
set.seed(123456)
plan(multisession, workers = 20L)
bmr <- benchmark(design,
store_models = TRUE,
store_backends = TRUE)
measures <- list (msr("classif.auc"),
msr("classif.acc"),
msr("classif.sensitivity"),
msr("classif.specificity"),
msr("classif.precision"),
msr("classif.fbeta"))
resample_perf <- as.data.table (bmr$score(measures = measures)) %>%
as.data.frame() %>%
dplyr::select (nr, task_id, learner_id, resampling_id, iteration, matches ("classif."))saveRDS(res_step, file="output/stepwise_performance.RDS")
# Save -------------------------------------------------------------------------
saveRDS(resample_perf,
"output/resample_perf.RDS")
saveRDS(bmr,
"output/resampling_models.RDS")saveRDS(res_step, file="output/stepwise_performance.RDS")
# Save -------------------------------------------------------------------------
saveRDS(resample_perf,
"output/resample_perf.RDS")
saveRDS(bmr,
"output/resampling_models.RDS")
# Evaluate results -------------------------------------------------------------
perf <- readRDS("output/resample_perf.RDS")
bmr <- readRDS("output/resampling_models.RDS")
perf_step <- readRDS("output/stepwise_performance.RDS")# Plot performance -------------------------------------------------------------
colz <- c("#030303", "#0000FF", "#EE0000", "#00CD00", "#EE9A00", "#912CEE", "#8B4726")
step_perf <- perf_step %>%
map ("metrics_test") %>%
bind_rows(.id = "task_id") %>%
mutate(task_id = ifelse (task_id == "np", "Neck pain",
ifelse (task_id == "ap", "Arm pain", "Disability"))) %>%
mutate (learner_id = "Stepwise") %>%
pivot_longer(cols = -c(task_id, learner_id),
names_to = "Metrics",
values_to = "Performance") %>%
mutate(Metrics = ifelse (Metrics == "auc", "AUC",
ifelse (Metrics == "acc", "ACC",
ifelse (Metrics == "sens", "Sens",
ifelse (Metrics == "spec", "Specs",
ifelse (Metrics == "precision", "Precs", "F1"))))))
perf.df <- perf %>%
dplyr::select (-c(resampling_id, iteration)) %>%
pivot_longer(cols = matches("classif"),
names_to = "Metrics",
values_to = "Performance") %>%
mutate(task_id = ifelse (task_id == "neckpain", "Neck pain",
ifelse (task_id == "armpain", "Arm pain", "Disability"))) %>%
mutate(learner_id = ifelse (learner_id == "scale.log", "Logistic",
ifelse (learner_id == "scale.encode.kknn.tuned", "Knn",
ifelse (learner_id == "scale.encode.xgb.tuned", "Xgb",
ifelse (learner_id == "scale.encode.lasso.tuned", "Lasso",
ifelse (learner_id == "scale.encode.rf.tuned", "RF",
ifelse (learner_id == "scale.encode.svm.tuned", "Svm", "ANN"))))))) %>%
mutate(Metrics = ifelse (Metrics == "classif.auc", "AUC",
ifelse (Metrics == "classif.acc", "ACC",
ifelse (Metrics == "classif.sensitivity", "Sens",
ifelse (Metrics == "classif.specificity", "Specs",
ifelse (Metrics == "classif.precision", "Precs", "F1")))))) %>%
bind_rows(step_perf) %>%
mutate (task_id = factor (task_id,
levels = c("Arm pain", "Neck pain", "Disability"))) %>%
mutate (Metrics = factor (Metrics,
levels = c("AUC", "ACC", "Precs", "F1", "Sens", "Specs"))) %>%
filter (learner_id != "Logistic") %>%
mutate (learner_id = factor (learner_id,
levels = c("Stepwise", "Knn",
"Lasso", "Xgb", "RF", "Svm", "ANN")))
pred_plot1 <- perf.df %>%
ggplot() +
geom_bar(aes (x = Metrics, y = Performance, fill = learner_id),
position = "dodge", stat = "identity") +
facet_wrap(~ task_id, ncol = 1) +
scale_fill_manual(values = colz) +
theme_cowplot() +
theme(axis.text.x = ggplot2::element_text(angle = 45, hjust = 1)) +
guides(fill=guide_legend(title="Learners"))
pred_plot1
ggsave (filename = "fig1.png",
plot = pred_plot1,
device = "png",
path = "manuscript",
width = 8,
height = 5,
units = "in",
dpi = 150)
perf.df <- perf.df %>%
mutate_if(is.numeric, round, 3)
perf.df %>%
filter (Metrics == "Specs") %>%
group_by(task_id) %>%
slice_max (Performance)
perf.df %>%
filter (Metrics == "AUC") %>%
group_by(task_id) %>%
slice_min (Performance)
# Extract hyperparameters selected ---------------------------------------------
tune_res <- extract_inner_tuning_results(bmr)
tune_res2 <- tune_res %>%
pivot_longer(cols = c("kknn.k": "svm.gamma"),
names_to = "hyperprms",
values_to = "params") %>%
dplyr::select (task_id, learner_id, hyperprms, params, classif.auc) %>%
na.omit() %>%
mutate(task_id = ifelse (task_id == "neckpain", "Neck pain",
ifelse (task_id == "armpain", "Arm pain", "Disability"))) %>%
mutate (task_id = factor (task_id,
levels = c("Arm pain", "Neck pain", "Disability"))) %>%
mutate(learner_id = ifelse (learner_id == "scale.log", "Logistic",
ifelse (learner_id == "scale.encode.kknn.tuned", "Knn",
ifelse (learner_id == "scale.encode.xgb.tuned", "Xgb",
ifelse (learner_id == "scale.encode.lasso.tuned", "Lasso",
ifelse (learner_id == "scale.encode.rf.tuned", "RF",
ifelse (learner_id == "scale.encode.svm.tuned", "Svm", "NNet"))))))) %>%
mutate (hyperprms = str_remove(hyperprms, ".*\\.")) %>%
mutate_if(is.numeric, round, 2)
# Export to word
my_path <- paste0("manuscript/table_",
"hyperparams",
".docx")
ft <- flextable(tune_res2) %>%
autofit()
my_doc <- read_docx() %>%
body_add_flextable(ft)
print (my_doc, target = my_path)
# Interpretable ML -------------------------------------------------------------
bmr2 <- as.data.table(bmr) %>%
mutate (Model = mlr3misc::map (learner, "model"))
####################### Arm pain ###############################################
df <- ap_dat
x <- df[which(names(df) != "outcome")]
y <- df$outcome
for (n in seq_along (bmr2$Model)) {
bmr2$iml_model[[n]] <- Predictor$new(bmr2$Model[[n]][[1]], data = x, y = y)
}
for (n in seq_along (bmr2$Model)[-c(1, 8, 15)]) {
bmr2$effect [[n]] <- FeatureImp$new(bmr2$iml_model[[n]], loss = "ce")
}
featimp <- list()
featimp_gg <- list()
for (n in seq_along (bmr2$Model)[-c(1, 8, 15)]) {
featimp [[n]] <- FeatureImp$new(bmr2$iml_model[[n]], loss = "ce")
}
for (n in c(2, 3, 5, 6, 7, 9, 10, 12, 13, 14, 16, 17, 19, 20, 21)) {
featimp_gg[[n]] <- featimp[[n]]$plot()
}
mody_gg <- function (x) {
x +
xlim (0.8, 2.5)
}
featimp_gg <- featimp_gg %>%
purrr::discard(is.null) %>%
purrr::map (~ .x +
xlim (0.8, 3) +
theme_cowplot() +
theme (axis.text.y = element_text(size = 10)))
pdf ("output/featimp.pdf", height = 20, width = 8)
plot_grid(plotlist = featimp_gg,
ncol = 2,
nrow = 8)
dev.off()
sessionInfo()