Package 'nestedcv'

Description

Produces a ggplot2 plot of stability (as SEM) of variable importance across models trained and tested across outer CV folds. Optionally overlays directionality for binary response or regression outcomes.

Usage

barplot_var_stability(
  x,
  final = TRUE,
  top = NULL,
  direction = 0,
  dir_labels = NULL,
  scheme = c("royalblue", "red"),
  breaks = NULL,
  percent = TRUE,
  level = 1,
  sort = TRUE
)

Arguments

Value

A ggplot2 plot

See Also

var_stability()

Description

Randomly samples predictors and averages the ranking to give an ensemble measure of predictor variable importance.

Usage

boot_filter(y, x, filterFUN, B = 50, nfilter = NULL, type = "index", ...)

Arguments

Value

Integer vector of indices of filtered parameters (type = "index") or if type = "full" a matrix of rankings from each bootstrap is returned.

See Also

boot_ttest()

Description

Randomly samples predictors and averages the ranking from filtering functions including ttest_filter(), wilcoxon_filter(), anova_filter(), correl_filter() and lm_filter() to give an ensemble measure of best predictors by repeated random sampling subjected to a statistical test.

Usage

boot_ttest(y, x, B = 50, ...)

boot_wilcoxon(y, x, B = 50, ...)

boot_anova(y, x, B = 50, ...)

boot_correl(y, x, B = 50, ...)

boot_lm(y, x, B = 50, ...)

Arguments

Value

Integer vector of indices of filtered parameters (type = "index"), or if type = "full", a matrix of rankings from each bootstrap is returned.

See Also

ttest_filter(), wilcoxon_filter(), anova_filter(), correl_filter(), lm_filter() and boot_filter()

Description

Filter using Boruta algorithm.

Usage

boruta_filter(
  y,
  x,
  select = c("Confirmed", "Tentative"),
  type = c("index", "names", "full"),
  ...
)

Arguments

Details

Boruta works differently from other filters in that it does not rank variables by variable importance, but tries to determine relevant features and divides features into Rejected, Tentative or Confirmed.

Value

Integer vector of indices of filtered parameters (type = "index") or character vector of names (type = "names") of filtered parameters. If type is "full" full output from Boruta is returned.

Description

Boxplots to show range of model predictors to identify exceptional predictors with excessively low or high values.

Usage

boxplot_expression(x, scheme = NULL, palette = "Dark 3", ...)

Arguments

Value

No return value

Author(s)

Myles Lewis

See Also

nestcv.glmnet

Description

Check class balance in training folds

Usage

class_balance(object)

## Default S3 method:
class_balance(object)

## S3 method for class 'nestcv.train'
class_balance(object)

Arguments

Value

Invisibly a table of the response classes in the training folds

Description

Extracts model coefficients from a fitted cva.glmnet() object.

Usage

## S3 method for class 'cva.glmnet'
coef(object, ...)

Arguments

Value

Sparse matrix containing coefficients from a cv.glmnet model

Description

Extracts coefficients from the final fit of a "nestcv.glmnet" object.

Usage

## S3 method for class 'nestcv.glmnet'
coef(object, s = object$final_param["lambda"], ...)

Arguments

Value

Vector or list of coefficients ordered with the intercept first, followed by highest absolute value to lowest.

Description

This function identifies predictors with r^2 above a given cut-off and produces an index of predictors to be removed. The function takes a matrix or data.frame of predictors, and the columns need to be ordered in terms of importance - first column of any pair that are correlated is retained and subsequent columns which correlate above the cut-off are flagged for removal.

Usage

collinear(x, rsq_cutoff = 0.9, rsq_method = "pearson", verbose = FALSE)

Arguments

Value

Integer vector of the indices of columns in x to remove due to collinearity

Description

Filter combining univariate (t-test or anova) filtering and reliefF filtering in equal measure.

Usage

combo_filter(y, x, nfilter, type = c("index", "names", "full"), ...)

Arguments

Value

Integer vector of indices of filtered parameters (type = "index") or character vector of names (type = "names") of filtered parameters. If type is "full" a list containing full outputs from either ttest_filter or anova_filter and relieff_filter is returned.

Description

Fast Pearson/Spearman correlation where y is vector, x is matrix, adapted from stats::cor.test.

Usage

correls2(y, x, method = "pearson", use = "complete.obs")

Arguments

Details

For speed, p-values for Spearman's test are computed by asymptotic t approximation, equivalent to cor.test with exact = FALSE.

Value

Matrix with columns containing the correlation statistic, either Pearson r or Spearman rho, and p-values for each column of x correlated against vector y

Description

Extracts coefficients from outer CV glmnet models from a nestcv.glmnet fitted object.

Usage

cv_coef(x, level = 1)

Arguments

Value

matrix of coefficients from outer CV glmnet models plus the final glmnet model. Coefficients for variables which are not present in a particular outer CV fold model are set to 0.

See Also

cv_varImp()

Description

Extracts variable importance or coefficients from outer CV glmnet models from a nestcv.train fitted object.

Usage

cv_varImp(x)

Arguments

Details

Note that caret::varImp() may require the model package to be fully loaded in order to function. During the fitting process caret often only loads the package by namespace.

Value

matrix of variable importance from outer CV fold caret models as well as the final model. Variable importance for variables which are not present in a particular outer CV fold model is set to 0.

See Also

cv_coef()

Description

Performs k-fold cross-validation for glmnet, including alpha mixing parameter.

Usage

cva.glmnet(x, y, nfolds = 10, alphaSet = seq(0.1, 1, 0.1), foldid = NULL, ...)

Arguments

Value

Object of S3 class "cva.glmnet", which is a list of the cv.glmnet objects for each value of alpha and alphaSet.

Author(s)

Myles Lewis

See Also

glmnet::cv.glmnet, glmnet::glmnet

Description

Convenience function for retrieving coefficients from a glmnet::cv.glmnet model at a specified lambda. Sparsity is removed and non-intercept coefficients are ranked by absolute value.

Usage

glmnet_coefs(fit, s, ...)

Arguments

Value

Vector or list of coefficients ordered with the intercept first, followed by highest absolute value to lowest.

Description

Filter using sparsity of elastic net regression using glmnet to calculate variable importance.

Usage

glmnet_filter(
  y,
  x,
  family = NULL,
  force_vars = NULL,
  nfilter = NULL,
  method = c("mean", "nonzero"),
  type = c("index", "names", "full"),
  ...
)

Arguments

Details

The glmnet elastic net mixing parameter alpha can be varied to include a larger number of predictors. Default alpha = 1 is pure LASSO, resulting in greatest sparsity, while alpha = 0 is pure ridge regression, retaining all predictors in the regression model. Note, the family argument is commonly needed, see glmnet::glmnet.

Value

Integer vector of indices of filtered parameters (type = "index") or character vector of names (type = "names") of filtered parameters. If type is "full" a named vector of variable importance is returned.

See Also

glmnet::glmnet

Description

Obtain predictions on held-out test inner CV folds

Usage

innercv_preds(x)

## S3 method for class 'nestcv.glmnet'
innercv_preds(x)

## S3 method for class 'nestcv.train'
innercv_preds(x)

Arguments

Value

Dataframe with columns testy and predy, and for binomial and multinomial models additional columns containing probabilities or log likelihood values.

Description

Build ROC (receiver operating characteristic) curve from left-out folds from inner CV. Object can be plotted using plot() or passed to functions pROC::auc() etc.

Usage

innercv_roc(x, direction = "<", ...)

Arguments

Value

"roc" object, see pROC::roc

Examples

## Example binary classification problem with P >> n
x <- matrix(rnorm(150 * 2e+04), 150, 2e+04)  # predictors
y <- factor(rbinom(150, 1, 0.5))  # binary response

## Partition data into 2/3 training set, 1/3 test set
trainSet <- caret::createDataPartition(y, p = 0.66, list = FALSE)

## t-test filter using whole dataset
filt <- ttest_filter(y, x, nfilter = 100)
filx <- x[, filt]

## Train glmnet on training set only using filtered predictor matrix
library(glmnet)
fit <- cv.glmnet(filx[trainSet, ], y[trainSet], family = "binomial")
plot(fit)

## Predict response on test partition
predy <- predict(fit, newx = filx[-trainSet, ], s = "lambda.min", type = "class")
predy <- as.vector(predy)
predyp <- predict(fit, newx = filx[-trainSet, ], s = "lambda.min", type = "response")
predyp <- as.vector(predyp)
output <- data.frame(testy = y[-trainSet], predy = predy, predyp = predyp)

## Results on test partition
## shows bias since univariate filtering was applied to whole dataset
predSummary(output)

## Nested CV
fit2 <- nestcv.glmnet(y, x, family = "binomial", alphaSet = 1,
                      filterFUN = ttest_filter,
                      filter_options = list(nfilter = 100),
                      n_outer_folds = 3)
summary(fit2)

## ROC plots
library(pROC)
testroc <- roc(output$testy, output$predyp, direction = "<")
inroc <- innercv_roc(fit2)
plot(fit2$roc)
lines(inroc, col = 'blue')
lines(testroc, col = 'red')
legend('bottomright', legend = c("Nested CV", "Left-out inner CV folds", 
                                 "Test partition, non-nested filtering"), 
       col = c("black", "blue", "red"), lty = 1, lwd = 2, bty = "n")

Description

Calculates performance metrics on inner CV held-out test folds: confusion matrix, accuracy and balanced accuracy for classification; ROC AUC for binary classification; RMSE, R^2 and mean absolute error (MAE) for regression.

Usage

innercv_summary(x)

Arguments

Value

Returns performance metrics from outer training folds, see predSummary.

See Also

predSummary

Examples

data(iris)
x <- iris[, 1:4]
y <- iris[, 5]

fit <- nestcv.glmnet(y, x,
                     family = "multinomial",
                     alpha = 1,
                     n_outer_folds = 3)
summary(fit)
innercv_summary(fit)

Description

Adds a precision-recall curve to a base graphics plot. It accepts an S3 object of class 'prc', see prc().

Usage

## S3 method for class 'prc'
lines(x, ...)

Arguments

Value

No return value

See Also

prc() plot.prc()

Description

Linear models are fitted on each predictor, with inclusion of variable names listed in force_vars in the model. Predictors are ranked by Akaike information criteria (AIC) value, or can be filtered by the p-value on the estimate of the coefficient for that predictor in its model.

Usage

lm_filter(
  y,
  x,
  force_vars = NULL,
  nfilter = NULL,
  p_cutoff = 0.05,
  rsq_cutoff = NULL,
  rsq_method = "pearson",
  type = c("index", "names", "full"),
  keep_factors = TRUE,
  method = 0L,
  mc.cores = 1
)

Arguments

Details

This filter is based on the model y ~ xvar + force_vars where y is the response vector, xvar are variables in columns taken sequentially from x and force_vars are optional covariates extracted from x. It uses RcppEigen::fastLmPure() with method = 0 as default since it is rank-revealing. method = 3 is significantly faster but can give errors in estimation of p-value with variables of zero variance. The algorithm attempts to detect these and set their stats to NA. NA in x are not tolerated.

Parallelisation is available via mclapply(). This is provided mainly for the use case of the filter being used as standalone. Nesting parallelisation inside of parallelised nestcv.glmnet() or nestcv.train() loops is not recommended.

Value

Integer vector of indices of filtered parameters (type = "index") or character vector of names (type = "names") of filtered parameters in order of linear model AIC. Any variables in force_vars which are incorporated into all models are listed first. If type = "full" a matrix of AIC value, sigma (residual standard error, see summary.lm), coefficient, t-statistic and p-value for each tested predictor is returned.

Description

Calculates Matthews correlation coefficient (MCC) which is in essence a correlation coefficient between the observed and predicted binary classifications. It has also been generalised to multi-class classification.

Usage

mcc(cm)

mcc_multi(cm)

Arguments

Details

Use mcc() for 2x2 tables (binary classification). mcc_multi() is for multi-class classification with k x k tables and is calculated using Gorodkin's method.

Value

Returns a value between -1 and +1. A coefficient of +1 represents a perfect prediction, 0 no better than random prediction and -1 indicates total disagreement between prediction and observation.

References

Gorodkin, J. (2004). Comparing two K-category assignments by a K-category correlation coefficient. Computational Biology and Chemistry. 28 (5): 367–374.

Description

Returns model metrics from nestedcv models. Extended metrics including

Usage

metrics(object, extra = FALSE, innerCV = FALSE, positive = 2)

Arguments

Details

Area under precision recall curve is estimated by trapezoidal estimation using MLmetrics::PRAUC().

For multi-class classification models, Matthews correlation coefficient is calculated using Gorodkin's method. Multi-class F1 score (macro F1) is calculated as the arithmetic mean of the class-wise F1 scores.

Value

A named numeric vector of performance metrics.

References

Gorodkin, J. (2004). Comparing two K-category assignments by a K-category correlation coefficient. Computational Biology and Chemistry. 28 (5): 367–374.

See Also

mcc()

Description

This function applies a cross-validation (CV) procedure for training Bayesian models with hierarchical shrinkage priors using the hsstan package. The function allows the option of embedded filtering of predictors for feature selection within the CV loop. Within each training fold, an optional filtering of predictors is performed, followed by fitting of an hsstsan model. Predictions on the testing folds are brought back together and error estimation/ accuracy determined. The default is 10-fold CV. The function is implemented within the nestedcv package. The hsstan models do not require tuning of meta-parameters and therefore only a single CV procedure is needed to evaluate performance. This is implemented using the outer CV procedure in the nestedcv package. Supports binary outcome (logistic regression) or continuous outcome. Multinomial models are currently not supported.

Usage

model.hsstan(y, x, unpenalized = NULL, ...)

Arguments

Details

Caution should be used when setting the number of cores available for parallelisation. The default setting in hsstan is to use 4 cores to parallelise the Markov chains of the Bayesian inference procedure. This can be switched off either by adding argument cores = 1 (passed on to rstan) or setting options(mc.cores = 1).

Argument cv.cores in outercv() controls parallelisation over the outer CV folds. On unix/mac setting cv.cores to >1 will induce nested parallelisation which will generate an error, unless parallelisation of the chains is disabled using cores = 1 or setting options(mc.cores = 1).

Nested parallelisation is feasible if cv.cores is >1 and multicore_fork = FALSE is set as this uses cluster based parallelisation instead. Beware that large numbers of processes will be spawned. If we are performing 10-fold cross-validation with 4 chains and set cv.cores = 10 then 40 processes will be invoked simultaneously.

Value

An object of class hsstan

Author(s)

Athina Spiliopoulou

See Also

outercv() hsstan::hsstan()

Examples

# Cross-validation is used to apply univariate filtering of predictors.
# only one CV split is needed (outercv) as the Bayesian model does not
# require learning of meta-parameters.

# control number of cores used for parallelisation over chains
oldopt <- options(mc.cores = 2)

# load iris dataset and simulate a continuous outcome
data(iris)
dt <- iris[, 1:4]
colnames(dt) <- c("marker1", "marker2", "marker3", "marker4")
dt <- as.data.frame(apply(dt, 2, scale))
dt$outcome.cont <- -3 + 0.5 * dt$marker1 + 2 * dt$marker2 + rnorm(nrow(dt), 0, 2)

library(hsstan)
# unpenalised covariates: always retain in the prediction model
uvars <- "marker1"
# penalised covariates: coefficients are drawn from hierarchical shrinkage
# prior
pvars <- c("marker2", "marker3", "marker4") # penalised covariates
# run cross-validation with univariate filter and hsstan
# dummy sampling for fast execution of example
# recommend 4 chains, warmup 1000, iter 2000 in practice
res.cv.hsstan <- outercv(y = dt$outcome.cont, x = dt[, c(uvars, pvars)],
                         model = "model.hsstan",
                         filterFUN = lm_filter,
                         filter_options = list(force_vars = uvars,
                                               nfilter = 2,
                                               p_cutoff = NULL,
                                               rsq_cutoff = 0.9),
                         n_outer_folds = 3,
                         chains = 2,
                         cv.cores = 1,
                         unpenalized = uvars, warmup = 100, iter = 200)
# view prediction performance based on testing folds
res.cv.hsstan$summary
# view coefficients for the final model
res.cv.hsstan$final_fit
# view covariates selected by the univariate filter
res.cv.hsstan$final_vars

# use hsstan package to examine the Bayesian model
sampler.stats(res.cv.hsstan$final_fit)
print(projsel(res.cv.hsstan$final_fit), digits = 4)  # adding marker2
options(oldopt)  # reset configuation

# Here adding `marker2` improves the model fit: substantial decrease of
# KL-divergence from the full model to the submodel. Adding `marker3` does
# not improve the model fit: no decrease of KL-divergence from the full model
# to the submodel.

Description

This function enables nested cross-validation (CV) with glmnet including tuning of elastic net alpha parameter. The function also allows the option of embedded filtering of predictors for feature selection nested within the outer loop of CV. Predictions on the outer test folds are brought back together and error estimation/ accuracy determined. The default is 10x10 nested CV.

Usage

nestcv.glmnet(
  y,
  x,
  family = c("gaussian", "binomial", "poisson", "multinomial", "cox", "mgaussian"),
  filterFUN = NULL,
  filter_options = NULL,
  balance = NULL,
  balance_options = NULL,
  modifyX = NULL,
  modifyX_useY = FALSE,
  modifyX_options = NULL,
  outer_method = c("cv", "LOOCV"),
  n_outer_folds = 10,
  n_inner_folds = 10,
  outer_folds = NULL,
  pass_outer_folds = FALSE,
  alphaSet = seq(0.1, 1, 0.1),
  min_1se = 0,
  keep = TRUE,
  outer_train_predict = FALSE,
  weights = NULL,
  penalty.factor = rep(1, ncol(x)),
  cv.cores = 1,
  finalCV = TRUE,
  na.option = "omit",
  verbose = FALSE,
  ...
)

Arguments

Details

glmnet does not tolerate missing values, so na.option = "omit" is the default.

Value

An object with S3 class "nestcv.glmnet"

Author(s)

Myles Lewis

Examples

## Example binary classification problem with P >> n
x <- matrix(rnorm(150 * 2e+04), 150, 2e+04)  # predictors
y <- factor(rbinom(150, 1, 0.5))  # binary response

## Partition data into 2/3 training set, 1/3 test set
trainSet <- caret::createDataPartition(y, p = 0.66, list = FALSE)

## t-test filter using whole dataset
filt <- ttest_filter(y, x, nfilter = 100)
filx <- x[, filt]

## Train glmnet on training set only using filtered predictor matrix
library(glmnet)
fit <- cv.glmnet(filx[trainSet, ], y[trainSet], family = "binomial")
plot(fit)

## Predict response on test partition
predy <- predict(fit, newx = filx[-trainSet, ], s = "lambda.min", type = "class")
predy <- as.vector(predy)
predyp <- predict(fit, newx = filx[-trainSet, ], s = "lambda.min", type = "response")
predyp <- as.vector(predyp)
output <- data.frame(testy = y[-trainSet], predy = predy, predyp = predyp)

## Results on test partition
## shows bias since univariate filtering was applied to whole dataset
predSummary(output)

## Nested CV
## n_outer_folds reduced to speed up example
fit2 <- nestcv.glmnet(y, x, family = "binomial", alphaSet = 1,
                      n_outer_folds = 3,
                      filterFUN = ttest_filter,
                      filter_options = list(nfilter = 100),
                      cv.cores = 2)
summary(fit2)
plot_lambdas(fit2, showLegend = "bottomright")

## ROC plots
library(pROC)
testroc <- roc(output$testy, output$predyp, direction = "<")
inroc <- innercv_roc(fit2)
plot(fit2$roc)
lines(inroc, col = 'blue')
lines(testroc, col = 'red')
legend('bottomright', legend = c("Nested CV", "Left-out inner CV folds", 
                                 "Test partition, non-nested filtering"), 
       col = c("black", "blue", "red"), lty = 1, lwd = 2, bty = "n")

Description

Provides a single loop of outer cross-validation to evaluate performance of ensemble models from SuperLearner package.

Usage

nestcv.SuperLearner(
  y,
  x,
  filterFUN = NULL,
  filter_options = NULL,
  weights = NULL,
  balance = NULL,
  balance_options = NULL,
  modifyX = NULL,
  modifyX_useY = FALSE,
  modifyX_options = NULL,
  outer_method = c("cv", "LOOCV"),
  n_outer_folds = 10,
  outer_folds = NULL,
  cv.cores = 1,
  final = TRUE,
  na.option = "pass",
  verbose = TRUE,
  ...
)

Arguments

Details

This performs an outer CV on SuperLearner package ensemble models to measure performance, allowing balancing of imbalanced datasets as well as filtering of predictors. SuperLearner prefers dataframes as inputs for the predictors. If x is a matrix it will be coerced to a dataframe and variable names adjusted by make.names().

Parallelisation of the outer CV folds is available on linux/mac, but not available on windows. On windows, snowSuperLearner() is called instead, so that parallelisation is performed across each call to SuperLearner.

Value

An object with S3 class "nestcv.SuperLearner"

Note

Care should be taken with some SuperLearner models e.g. SL.gbm as some models have multicore enabled by default, which can lead to huge numbers of processes being spawned.

See Also

SuperLearner::SuperLearner()

Description

This function applies nested cross-validation (CV) to training of models using the caret package. The function also allows the option of embedded filtering of predictors for feature selection nested within the outer loop of CV. Predictions on the outer test folds are brought back together and error estimation/ accuracy determined. The default is 10x10 nested CV.

Usage

nestcv.train(
  y,
  x,
  method = "rf",
  filterFUN = NULL,
  filter_options = NULL,
  weights = NULL,
  balance = NULL,
  balance_options = NULL,
  modifyX = NULL,
  modifyX_useY = FALSE,
  modifyX_options = NULL,
  outer_method = c("cv", "LOOCV"),
  n_outer_folds = 10,
  n_inner_folds = 10,
  outer_folds = NULL,
  inner_folds = NULL,
  pass_outer_folds = FALSE,
  cv.cores = 1,
  multicore_fork = (Sys.info()["sysname"] != "Windows"),
  metric = ifelse(is.factor(y), "logLoss", "RMSE"),
  trControl = NULL,
  tuneGrid = NULL,
  savePredictions = "final",
  outer_train_predict = FALSE,
  finalCV = TRUE,
  na.option = "pass",
  verbose = TRUE,
  ...
)

Arguments

Details

When finalCV = TRUE, the final fit on the whole data using is performed first. This helps flag errors generated by caret such as missing packages. Parallelisation of the final fit when finalCV = TRUE is performed in caret using registerDoParallel. caret itself uses foreach.

Parallelisation is performed on the outer CV folds using parallel::mclapply by default on unix/mac and parallel::parLapply on windows. mclapply uses forking which is faster. But some models use multi-threading which may cause issues in some circumstances with forked multicore processing. Setting multicore_fork to FALSE is slower but can alleviate some caret errors.

If the outer folds are run using parallelisation, then parallelisation in caret must be off, otherwise an error will be generated. Alternatively if you wish to use parallelisation in caret, then parallelisation in nestcv.train can be fully disabled by leaving cv.cores = 1.

xgboost models fitted via caret using method = "xgbTree" or "xgbLinear" invoke openMP multithreading on linux/windows by default which causes nestcv.train to fail when cv.cores >1 (nested parallelisation). Mac OS is unaffected. In order to prevent this, nestcv.train() sets openMP threads to 1 if cv.cores >1.

For classification, metric defaults to using 'logLoss' with the trControl arguments ⁠classProbs = TRUE, summaryFunction = mnLogLoss⁠, rather than 'Accuracy' which is the default classification metric in caret. See caret::trainControl(). LogLoss is arguably more consistent than Accuracy for tuning parameters in datasets with small sample size.

Models can be fitted with a single set of fixed parameters, in which case trControl defaults to trainControl(method = "none") which disables inner CV as it is unnecessary. See https://topepo.github.io/caret/model-training-and-tuning.html#fitting-models-without-parameter-tuning

Value

An object with S3 class "nestcv.train"

Author(s)

Myles Lewis

Examples

## sigmoid function
sigmoid <- function(x) {1 / (1 + exp(-x))}

## load iris dataset and simulate a binary outcome
data(iris)
x <- iris[, 1:4]
colnames(x) <- c("marker1", "marker2", "marker3", "marker4")
x <- as.data.frame(apply(x, 2, scale))
y2 <- sigmoid(0.5 * x$marker1 + 2 * x$marker2) > runif(nrow(x))
y2 <- factor(y2, labels = c("class1", "class2"))

## Example using random forest with caret
cvrf <- nestcv.train(y2, x, method = "rf",
                     n_outer_folds = 3,
                     cv.cores = 2)
summary(cvrf)

## Example of glmnet tuned using caret
## set up small tuning grid for quick execution
## length.out of 20-100 is usually recommended for lambda
## and more alpha values ranging from 0-1
tg <- expand.grid(lambda = exp(seq(log(2e-3), log(1e0), length.out = 5)),
                  alpha = 1)

ncv <- nestcv.train(y = y2, x = x,
                    method = "glmnet",
                    n_outer_folds = 3,
                    tuneGrid = tg, cv.cores = 2)
summary(ncv)

## plot tuning for outer fold #1
plot(ncv$outer_result[[1]]$fit, xTrans = log)

## plot final ROC curve
plot(ncv$roc)

## plot ROC for left-out inner folds
inroc <- innercv_roc(ncv)
plot(inroc)

## example to show use of custom fold indices for 5 x 5-fold nested CV
library(caret)
y <- iris$Species
out_folds <- createFolds(y, k = 5)
in_folds <- lapply(out_folds, function(i) {
  ytrain <- y[-i]
  createFolds(ytrain, k = 5)
})

res <- nestcv.train(y, x, method="rf", cv.cores = 2,
                    pass_outer_folds = TRUE,
                    inner_folds = in_folds,
                    outer_folds = out_folds)
summary(res)
res$outer_folds
res$final_fit$control$indexOut  # same as outer_folds

Description

Fast one-hot encoding of all factor and character columns in a dataframe to convert it into a numeric matrix by creating dummy (binary) columns.

Usage

one_hot(x, all_levels = FALSE, rename_binary = TRUE, sep = ".")

Arguments

Details

Binary factor columns and logical columns are converted to integers (0 or 1). Multi-level unordered factors are converted to multiple columns of 0/1 (dummy variables): if all_levels is set to FALSE (the default), then the first level is assumed to be a reference level and additional columns are created for each additional level; if all_levels is set to TRUE one column is used for each level. Unused levels are dropped. Character columns are first converted to factors and then encoded. Ordered factors are replaced by their internal codes. Numeric or integer columns are left untouched.

Having dummy variables for all levels of a factor can cause problems with multicollinearity in regression (the dummy variable trap), so all_levels is set to FALSE by default which is necessary for regression models such as glmnet (equivalent to full rank parameterisation). However, setting all_levels to TRUE can aid with interpretability (e.g. with SHAP values), and in some cases filtering might result in some dummy variables being excluded. Note this function is designed to quickly generate dummy variables for more general machine learning purposes. To create a proper design matrix object for regression models, use model.matrix().

Value

A numeric matrix with the same number of rows as the input data. Dummy variable columns replace the input factor or character columns. Numeric columns are left intact.

See Also

caret::dummyVars(), model.matrix()

Examples

data(iris)
x <- iris
x2 <- one_hot(x)
head(x2)  # 3 columns for Species

x2 <- one_hot(x, all_levels = FALSE)
head(x2)  # 2 columns for Species

Description

This is a convenience function designed to use a single loop of cross-validation to quickly evaluate performance of specific models (random forest, naive Bayes, lm, glm) with fixed hyperparameters and no tuning. If tuning of parameters on data is required, full nested CV with inner CV is needed to tune model hyperparameters (see nestcv.train).

Usage

outercv(y, ...)

## Default S3 method:
outercv(
  y,
  x,
  model,
  filterFUN = NULL,
  filter_options = NULL,
  weights = NULL,
  balance = NULL,
  balance_options = NULL,
  modifyX = NULL,
  modifyX_useY = FALSE,
  modifyX_options = NULL,
  outer_method = c("cv", "LOOCV"),
  n_outer_folds = 10,
  outer_folds = NULL,
  cv.cores = 1,
  multicore_fork = (Sys.info()["sysname"] != "Windows"),
  predict_type = "prob",
  outer_train_predict = FALSE,
  returnList = FALSE,
  final = TRUE,
  na.option = "pass",
  verbose = FALSE,
  suppressMsg = verbose,
  ...
)

## S3 method for class 'formula'
outercv(
  formula,
  data,
  model,
  outer_method = c("cv", "LOOCV"),
  n_outer_folds = 10,
  outer_folds = NULL,
  cv.cores = 1,
  multicore_fork = (Sys.info()["sysname"] != "Windows"),
  predict_type = "prob",
  outer_train_predict = FALSE,
  verbose = FALSE,
  suppressMsg = verbose,
  ...,
  na.action = na.fail
)

Arguments

Details

Some predictive model functions do not have an x & y interface. If the function specified by model requires a formula, x & y will be merged into a dataframe with model() called with a formula equivalent to y ~ ..

The S3 formula method for outercv is not really recommended with large data sets - it is envisaged to be primarily used to compare performance of more basic models e.g. lm() specified by formulae for example incorporating interactions. NOTE: filtering is not available if outercv is called with a formula - use the x-y interface instead.

An alternative method of tuning a single model with fixed parameters is to use nestcv.train with tuneGrid set as a single row of a data.frame. The parameters which are needed for a specific model can be identified using caret::modelLookup().

Case weights can be passed to model function which accept these, however outercv assumes that these are passed to the model via an argument named weights.

Note that in the case of model = "lm", although additional arguments e.g. subset, weights, offset are passed into the model function via "..." the scoping is known to go awry. Avoid using these arguments with model = "lm".

NA handling differs between the default S3 method and the formula S3 method. The na.option argument takes a character string, while the more typical na.action argument takes a function.

Value

An object with S3 class "outercv"

Examples

## Classification example

## sigmoid function
sigmoid <- function(x) {1 / (1 + exp(-x))}

# load iris dataset and simulate a binary outcome
data(iris)
dt <- iris[, 1:4]
colnames(dt) <- c("marker1", "marker2", "marker3", "marker4")
dt <- as.data.frame(apply(dt, 2, scale))
x <- dt
y2 <- sigmoid(0.5 * dt$marker1 + 2 * dt$marker2) > runif(nrow(dt))
y2 <- factor(y2)

## Random forest
library(randomForest)
cvfit <- outercv(y2, x, "randomForest")
summary(cvfit)
plot(cvfit$roc)

## Mixture discriminant analysis (MDA)
if (requireNamespace("mda", quietly = TRUE)) {
  library(mda)
  cvfit <- outercv(y2, x, "mda", predict_type = "posterior")
  summary(cvfit)
}


## Example with continuous outcome
y <- -3 + 0.5 * dt$marker1 + 2 * dt$marker2 + rnorm(nrow(dt), 0, 2)
dt$outcome <- y

## simple linear model - formula interface
cvfit <- outercv(outcome ~ ., data = dt, model = "lm")
summary(cvfit)

## random forest for regression
cvfit <- outercv(y, x, "randomForest")
summary(cvfit)

## example with lm_filter() to reduce input predictors
cvfit <- outercv(y, x, "randomForest", filterFUN = lm_filter,
                 filter_options = list(nfilter = 2, p_cutoff = NULL))
summary(cvfit)

Description

Plot of cross-validated glmnet alpha parameter against deviance for each outer CV fold.

Usage

plot_alphas(x, col = NULL, ...)

Arguments

Value

No return value

Author(s)

Myles Lewis

See Also

nestcv.glmnet

Description

Plots the main tuning parameter in models built using caret::train

Usage

plot_caret(x, error.col = "darkgrey", ...)

Arguments

Value

No return value

Description

Plot of cross-validated glmnet lambda parameter against deviance for each outer CV fold.

Usage

plot_lambdas(
  x,
  scheme = NULL,
  palette = "Dark 3",
  showLegend = if (x$outer_method == "cv") "topright" else NULL,
  ...
)

Arguments

Value

No return value

Author(s)

Myles Lewis

See Also

nestcv.glmnet

Description

SHAP importance bar plot

Usage

plot_shap_bar(
  shap,
  x,
  sort = TRUE,
  labels = c("Negative", "Positive"),
  top = NULL
)

Arguments

Value

A ggplot2 plot

Description

SHAP importance beeswarm plot

Usage

plot_shap_beeswarm(
  shap,
  x,
  cex = 0.25,
  corral = "random",
  corral.width = 0.7,
  scheme = c("deepskyblue2", "purple3", "red"),
  sort = TRUE,
  top = NULL,
  ...
)

Arguments

Value

A ggplot2 plot

Description

Plots variables selected in models ranked by variable importance across the outer folds as well as the final model.

Usage

plot_var_ranks(
  x,
  sort = TRUE,
  scheme = NULL,
  cex = 1,
  corral.width = 0.75,
  ...
)

hist_var_ranks(x, sort = TRUE, scheme = NULL)

Arguments

Value

A ggplot2 plot.

Description

Produces a ggplot2 plot of stability (as SEM) of variable importance across models trained and tested across outer CV folds. Overlays frequency with which variables are selected across the outer folds and optionally overlays directionality for binary response outcome.

Usage

plot_var_stability(
  x,
  final = TRUE,
  top = NULL,
  direction = 0,
  dir_labels = NULL,
  scheme = c("royalblue", "red"),
  breaks = NULL,
  percent = TRUE,
  level = 1,
  sort = TRUE
)

Arguments

Value

A ggplot2 plot

See Also

var_stability()

Description

Plot of variable importance of coefficients of a final fitted 'nestedcv.glmnet' model using ggplot2. Mean expression can be overlaid as the size of points as this can be informative in models of biological attributes.

Usage

plot_varImp(x, abs = TRUE, size = TRUE)

Arguments

Value

Returns a ggplot2 plot

Description

Different types of plot showing cross-validated tuning of alpha and lambda from elastic net regression via glmnet::glmnet. If xaxis is set to "lambda", log lambda is on the x axis while the tuning metric (log loss, deviance, accuracy, AUC etc) is on the y axis. Multiple alpha values are shown by different colours. If xaxis is set to "alpha", alpha is on the x axis with the tuning metric on y, with error bars showing metric SD. if xaxis is set to "nvar" the number of non-zero coefficients is shown on x and how this relates to model deviance/ accuracy on y.

Usage

## S3 method for class 'cva.glmnet'
plot(
  x,
  xaxis = c("lambda", "alpha", "nvar"),
  errorBar = (xaxis == "alpha"),
  errorWidth = 0.015,
  min.pch = NULL,
  scheme = NULL,
  palette = "zissou",
  showLegend = "bottomright",
  ...
)

Arguments

Value

No return value

Author(s)

Myles Lewis

See Also

nestcv.glmnet

Description

Plots a precision-recall curve using base graphics. It accepts an S3 object of class 'prc', see prc().

Usage

## S3 method for class 'prc'
plot(x, ...)

Arguments

Value

No return value

See Also

prc()

Examples

library(mlbench)
data(Sonar)
y <- Sonar$Class
x <- Sonar[, -61]

fit1 <- nestcv.glmnet(y, x, family = "binomial", alphaSet = 1, cv.cores = 2)
fit1$prc <- prc(fit1)  # calculate precision-recall curve

fit2 <- nestcv.train(y, x, method = "gbm", cv.cores = 2)
fit2$prc <- prc(fit2)

plot(fit1$prc)
lines(fit2$prc, col = "red")

Description

Filter using coefficients from partial least squares (PLS) regression to select optimal predictors.

Usage

pls_filter(
  y,
  x,
  force_vars = NULL,
  nfilter,
  ncomp = 5,
  scale_x = TRUE,
  type = c("index", "names", "full"),
  ...
)

Arguments

Details

The best predictors may overlap between components, so if nfilter is specified as a vector, the total number of unique predictors returned may be variable.

Value

Integer vector of indices of filtered parameters (type = "index") or character vector of names (type = "names") of filtered parameters. If type is "full" full output of coefficients from plsr is returned as a list for each model component ordered by highest absolute coefficient.

Description

Builds a precision-recall curve for a 'nestedcv' model using prediction() and performance() functions from the ROCR package and returns an object of class 'prc' for plotting.

Usage

prc(...)

## Default S3 method:
prc(response, predictor, positive = 2, ...)

## S3 method for class 'data.frame'
prc(output, ...)

## S3 method for class 'nestcv.glmnet'
prc(object, ...)

## S3 method for class 'nestcv.train'
prc(object, ...)

## S3 method for class 'nestcv.SuperLearner'
prc(object, ...)

## S3 method for class 'outercv'
prc(object, ...)

## S3 method for class 'repeatcv'
prc(object, ...)

Arguments

Value

An object of S3 class 'prc' containing the following fields:

Examples

library(mlbench)
data(Sonar)
y <- Sonar$Class
x <- Sonar[, -61]

fit1 <- nestcv.glmnet(y, x, family = "binomial", alphaSet = 1, cv.cores = 2)

fit1$prc <- prc(fit1)  # calculate precision-recall curve
fit1$prc$auc  # precision-recall AUC value

fit2 <- nestcv.train(y, x, method = "gbm", cv.cores = 2)
fit2$prc <- prc(fit2)
fit2$prc$auc

plot(fit1$prc, ylim = c(0, 1))
lines(fit2$prc, col = "red")

res <- nestcv.glmnet(y, x, family = "binomial", alphaSet = 1) |>
  repeatcv(n = 4, rep.cores = 2)

res$prc <- prc(res)  # precision-recall curve on repeated predictions
plot(res$prc)

Description

Prediction wrapper functions to enable the use of the fastshap package for generating SHAP values from nestedcv trained models.

Usage

pred_nestcv_glmnet(x, newdata)

pred_nestcv_glmnet_class1(x, newdata)

pred_nestcv_glmnet_class2(x, newdata)

pred_nestcv_glmnet_class3(x, newdata)

pred_train(x, newdata)

pred_train_class1(x, newdata)

pred_train_class2(x, newdata)

pred_train_class3(x, newdata)

pred_SuperLearner(x, newdata)

Arguments

Details

These prediction wrapper functions are designed to be used with the fastshap package. The functions pred_nestcv_glmnet and pred_train work for nestcv.glmnet and nestcv.train models respectively for either binary classification or regression.

For multiclass classification use pred_nestcv_glmnet_class1, 2 and 3 for the first 3 classes. Similarly pred_train_class1 etc for nestcv.train objects. These functions can be inspected and easily modified to analyse further classes.

Value

prediction wrapper function designed for use with fastshap::explain()

Examples

library(fastshap)

# Boston housing dataset
library(mlbench)
data(BostonHousing2)
dat <- BostonHousing2
y <- dat$cmedv
x <- subset(dat, select = -c(cmedv, medv, town, chas))

# Fit a glmnet model using nested CV
# Only 3 outer CV folds and 1 alpha value for speed
fit <- nestcv.glmnet(y, x, family = "gaussian", n_outer_folds = 3, alphaSet = 1)

# Generate SHAP values using fastshap::explain
# Only using 5 repeats here for speed, but recommend higher values of nsim
sh <- explain(fit, X=x, pred_wrapper = pred_nestcv_glmnet, nsim = 1)

# Plot overall variable importance
plot_shap_bar(sh, x)

# Plot beeswarm plot
plot_shap_beeswarm(sh, x, size = 1)

Description

Makes predictions from a cross-validated glmnet model with optimal value of lambda and alpha.

Usage

## S3 method for class 'cva.glmnet'
predict(object, newx, s = "lambda.1se", ...)

Arguments

Value

Object returned depends on arguments in ... such as type.

Description

Draws from the posterior predictive distribution of the outcome.

Usage

## S3 method for class 'hsstan'
predict(object, newdata = NULL, type = NULL, ...)

Arguments

Value

For a binary outcome and type = NULL, a character vector with the name of the class that has the highest probability for each sample. For a binary outcome and type = prob, a 2-dimensional matrix with the probability of class 0 and of class 1 for each sample. For a continuous outcome a numeric vector with the predicted value for each sample.

Author(s)

Athina Spiliopoulou

Description

Obtains predictions from the final fitted model from a nestcv.glmnet object.

Usage

## S3 method for class 'nestcv.glmnet'
predict(object, newdata, s = object$final_param["lambda"], modify = FALSE, ...)

Arguments

Details

Checks for missing predictors and if these are sparse (i.e. have zero coefficients) columns of 0 are automatically added to enable prediction to proceed.

Value

Object returned depends on the ... argument passed to predict method for glmnet objects.

See Also

glmnet::glmnet

Description

Quick function to calculate performance metrics: confusion matrix, accuracy and balanced accuracy for classification; ROC AUC for binary classification; RMSE, R^2 and MAE for regression. Multi-class AUC is returned for multinomial classification.

Usage

predSummary(output, family = "")

Arguments

Details

For multinomial classification, multi-class AUC as defined by Hand and Till is calculated using pROC::multiclass.roc().

Multi-class balanced accuracy is calculated as the mean of the Recall for each class.

R^2 (coefficient of determination) is calculated as 1 - rss / tss, where rss = residual sum of squares, tss = total sum of squares. Pearson r^2 is also provided. Pearson r^2 can only range from 0 to 1, whereas R^2 can range from 1 to -Inf.

Value

An object of class 'predSummary'. For classification a list is returned containing the confusion matrix table and a vector containing accuracy and balanced accuracy for classification, ROC AUC for classification. For regression a vector containing RMSE, R^2 and MAE is returned. For glmnet 'cox' models, Harrell's C-index is returned.

For glmnet 'mgaussian' models, an object of class 'predSummaryMulti' is returned which is a list of vectors with regression metrics (RMSE, R^2, MAE) for each response variable (i.e. each y column).

See Also

metrics()

Description

Random oversampling of the minority group(s) or undersampling of the majority group to compensate for class imbalance in datasets.

Usage

randomsample(y, x, minor = NULL, major = 1, yminor = NULL)

Arguments

Details

minor < 1 and major > 1 are ignored.

Value

List containing extended matrix x of synthesised data and extended response vector y

Examples

## Imbalanced dataset
set.seed(1, "L'Ecuyer-CMRG")
x <- matrix(rnorm(150 * 2e+04), 150, 2e+04)  #' predictors
y <- factor(rbinom(150, 1, 0.2))  #' imbalanced binary response
table(y)

## first 30 parameters are weak predictors
x[, 1:30] <- rnorm(150 * 30, 0, 1) + as.numeric(y)*0.5

## Balance x & y outside of CV loop by random oversampling minority group
out <- randomsample(y, x)
y2 <- out$y
x2 <- out$x
table(y2)

## Nested CV glmnet with unnested balancing by random oversampling on
## whole dataset
fit1 <- nestcv.glmnet(y2, x2, family = "binomial", alphaSet = 1,
                      cv.cores=2,
                      filterFUN = ttest_filter)
fit1$summary

## Balance x & y outside of CV loop by random oversampling minority group
out <- randomsample(y, x, minor=1, major=0.4)
y2 <- out$y
x2 <- out$x
table(y2)

## Nested CV glmnet with unnested balancing by random undersampling on
## whole dataset
fit1b <- nestcv.glmnet(y2, x2, family = "binomial", alphaSet = 1,
                       cv.cores=2,
                       filterFUN = ttest_filter)
fit1b$summary

## Balance x & y outside of CV loop by SMOTE
out <- smote(y, x)
y2 <- out$y
x2 <- out$x
table(y2)

## Nested CV glmnet with unnested balancing by SMOTE on whole dataset
fit2 <- nestcv.glmnet(y2, x2, family = "binomial", alphaSet = 1,
                      cv.cores=2,
                      filterFUN = ttest_filter)
fit2$summary

## Nested CV glmnet with nested balancing by random oversampling
fit3 <- nestcv.glmnet(y, x, family = "binomial", alphaSet = 1,
                      cv.cores=2,
                      balance = "randomsample",
                      filterFUN = ttest_filter)
fit3$summary
class_balance(fit3)

## Plot ROC curves
plot(fit1$roc, col='green')
lines(fit1b$roc, col='red')
lines(fit2$roc, col='blue')
lines(fit3$roc)
legend('bottomright', legend = c("Unnested random oversampling", 
                                 "Unnested SMOTE",
                                 "Unnested random undersampling",
                                 "Nested balancing"), 
       col = c("green", "blue", "red", "black"), lty=1, lwd=2)

Description

Fits a random forest model via the ranger package and ranks variables by variable importance.

Usage

ranger_filter(
  y,
  x,
  nfilter = NULL,
  type = c("index", "names", "full"),
  num.trees = 1000,
  mtry = ncol(x) * 0.2,
  ...
)

Arguments

Details

This filter uses the ranger() function from the ranger package. Variable importance is calculated using mean decrease in gini impurity.

Value

Integer vector of indices of filtered parameters (type = "index") or character vector of names (type = "names") of filtered parameters. If type is "full" a named vector of variable importance is returned.

Description

Uses ReliefF algorithm from the CORElearn package to rank predictors in order of importance.

Usage

relieff_filter(
  y,
  x,
  nfilter = NULL,
  estimator = "ReliefFequalK",
  type = c("index", "names", "full"),
  ...
)

Arguments

Value

Integer vector of indices of filtered parameters (type = "index") or character vector of names (type = "names") of filtered parameters. If type is "full" a named vector of variable importance is returned.

See Also

CORElearn::attrEval()

Description

Performs repeated calls to a nestedcv model to determine performance across repeated runs of nested CV.

Usage

repeatcv(
  expr,
  n = 5,
  repeat_folds = NULL,
  keep = FALSE,
  extra = FALSE,
  progress = TRUE,
  rep.cores = 1L
)

Arguments

Details

We recommend using this with the R pipe ⁠|>⁠ (see examples).

When comparing models, it is recommended to fix the sets of outer CV folds used across each repeat for comparing performance between models. The function repeatfolds() can be used to create a fixed set of outer CV folds for each repeat.

Parallelisation over repeats is performed using parallel::mclapply (not available on windows). Beware that cv.cores can still be set within calls to nestedcv models (= nested parallelisation). This means that rep.cores x cv.cores number of processes/forks will be spawned, so be careful not to overload your CPU. In general parallelisation of repeats using rep.cores is faster than parallelisation using cv.cores.

Value

List of S3 class 'repeatcv' containing:

Examples

data("iris")
dat <- iris
y <- dat$Species
x <- dat[, 1:4]

res <- nestcv.glmnet(y, x, family = "multinomial", alphaSet = 1,
                     n_outer_folds = 4) |>
       repeatcv(3, rep.cores = 2)
res
summary(res)

## set up fixed fold indices
set.seed(123, "L'Ecuyer-CMRG")
folds <- repeatfolds(y, repeats = 3, n_outer_folds = 4)
res <- nestcv.glmnet(y, x, family = "multinomial", alphaSet = 1,
                     n_outer_folds = 4) |>
       repeatcv(3, repeat_folds = folds, rep.cores = 2)
res

Description

Create folds for repeated nested CV

Usage

repeatfolds(y, repeats = 5, n_outer_folds = 10)

Arguments

Value

List containing indices of outer CV folds

Examples

data("iris")
dat <- iris
y <- dat$Species
x <- dat[, 1:4]

## set up fixed fold indices
set.seed(123, "L'Ecuyer-CMRG")
folds <- repeatfolds(y, repeats = 3, n_outer_folds = 4)

res <- nestcv.glmnet(y, x, family = "multinomial", alphaSet = 1,
                     n_outer_folds = 4, cv.cores = 2) |>
       repeatcv(3, repeat_folds = folds)
res

Description

Fits a random forest model and ranks variables by variable importance.

Usage

rf_filter(
  y,
  x,
  nfilter = NULL,
  type = c("index", "names", "full"),
  ntree = 1000,
  mtry = ncol(x) * 0.2,
  ...
)

Arguments

Details

This filter uses the randomForest() function from the randomForest package. Variable importance is calculated using the randomForest::importance function, specifying type 1 = mean decrease in accuracy. See randomForest::importance.

Value

Integer vector of indices of filtered parameters (type = "index") or character vector of names (type = "names") of filtered parameters. If type is "full" a named vector of variable importance is returned.

Description

Slims nestedcv objects to only the models in the outer CV folds.

Usage

slim(x)

Arguments

Value

For 'nestedcv' objects, a list object of the same class but only containing outer_result. For 'cva.glmnet' models, only the cv.glmnet model with the best alpha value is kept. Models for all other values of alpha are discarded.

See Also

nestcv.glmnet() nestcv.train() outercv() cva.glmnet()

Description

Synthetic Minority Oversampling Technique (SMOTE) algorithm for imbalanced classification data.

Usage

smote(y, x, k = 5, over = NULL, yminor = NULL)

Arguments

Value

List containing extended matrix x of synthesised data and extended response vector y

References

Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer, W. P. (2002). Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16:321-357.

Description

Univariate statistic filter for dataframes of predictors with mixed numeric and categorical datatypes. Different statistical tests are used depending on the data type of response vector and predictors:

Binary class response: bin_stat_filter()

t-test for continuous data, chi-squared test for categorical data

Multiclass response: class_stat_filter()

one-way ANOVA for continuous data, chi-squared test for categorical data

Continuous response: cor_stat_filter()

correlation (or linear regression) for continuous data and binary data, one-way ANOVA for categorical data

Usage

stat_filter(y, x, ...)

bin_stat_filter(
  y,
  x,
  force_vars = NULL,
  nfilter = NULL,
  p_cutoff = 0.05,
  rsq_cutoff = NULL,
  type = c("index", "names", "full", "list"),
  ...
)

class_stat_filter(
  y,
  x,
  force_vars = NULL,
  nfilter = NULL,
  p_cutoff = 0.05,
  rsq_cutoff = NULL,
  type = c("index", "names", "full", "list"),
  ...
)

cor_stat_filter(
  y,
  x,
  cor_method = c("pearson", "spearman", "lm"),
  force_vars = NULL,
  nfilter = NULL,
  p_cutoff = 0.05,
  rsq_cutoff = NULL,
  rsq_method = "pearson",
  type = c("index", "names", "full", "list"),
  ...
)

Arguments

Details

stat_filter() is a wrapper which calls bin_stat_filter(), class_stat_filter() or cor_stat_filter() depending on whether y is binary, multiclass or continuous respectively. Ordered factors are converted to numeric (integer) levels and analysed as if continuous.

Value

Integer vector of indices of filtered parameters (type = "index") or character vector of names (type = "names") of filtered parameters in order of test p-value. If type is "full" full output is returned containing a dataframe of statistical results. If type is "list" the output is returned as a list of 2 matrices containing statistical results separated by continuous and categorical predictors.

Examples

library(mlbench)
data(BostonHousing2)
dat <- BostonHousing2
y <- dat$cmedv  ## continuous outcome
x <- subset(dat, select = -c(cmedv, medv, town))

stat_filter(y, x, type = "full")
stat_filter(y, x, nfilter = 5, type = "names")
stat_filter(y, x)

data(iris)
y <- iris$Species  ## 3 class outcome
x <- subset(iris, select = -Species)
stat_filter(y, x, type = "full")

Description

Summarise variables

Usage

summary_vars(x)

Arguments

Value

A matrix with variables in rows and mean, median and SD for each variable or number of levels if the variable is a factor. If NA are detected, an extra column n.NA is added with the numbers of NA for each variable.

Description

Performs supervised principle component analysis (PCA) after filtering dataset to help determine whether filtering has been useful for separating samples according to the outcome variable.

Usage

supervisedPCA(y, x, filterFUN = NULL, filter_options = NULL, plot = TRUE, ...)

Arguments

Value

If plot=TRUE returns a ggplot2 plot, otherwise returns the principle component scores.

Description

Obtain predictions on outer training folds which can be used for performance metrics and ROC curves.

Usage

train_preds(x)

Arguments

Details

Note: the argument outer_train_predict must be set to TRUE in the original call to either nestcv.glmnet, nestcv.train or outercv.

Value

Dataframe with columns ytrain and predy containing observed and predicted values from training folds. For binomial and multinomial models additional columns are added with class probabilities or log likelihood values.

Description

Build ROC (receiver operating characteristic) curve from outer training folds. Object can be plotted using plot() or passed to functions pROC::auc() etc.

Usage

train_roc(x, direction = "<", ...)

Arguments

Details

Note: the argument outer_train_predict must be set to TRUE in the original call to either nestcv.glmnet, nestcv.train or outercv.

Value

"roc" object, see pROC::roc

Description

Calculates performance metrics on outer training folds: confusion matrix, accuracy and balanced accuracy for classification; ROC AUC for binary classification; RMSE, R^2 and mean absolute error (MAE) for regression.

Usage

train_summary(x)

Arguments

Details

Note: the argument outer_train_predict must be set to TRUE in the original call to either nestcv.glmnet, nestcv.train or outercv.

Value

Returns performance metrics from outer training folds, see predSummary

See Also

predSummary

Examples

data(iris)
x <- iris[, 1:4]
y <- iris[, 5]

fit <- nestcv.glmnet(y, x,
                     family = "multinomial",
                     alpha = 1,
                     outer_train_predict = TRUE,
                     n_outer_folds = 3)
summary(fit)
innercv_summary(fit)
train_summary(fit)

fit2 <- nestcv.train(y, x,
                    model="svm",
                    outer_train_predict = TRUE,
                    n_outer_folds = 3,
                    cv.cores = 2)
summary(fit2)
innercv_summary(fit2)
train_summary(fit2)

Description

A selection of simple univariate filters using t-test, Wilcoxon test, one-way ANOVA or correlation (Pearson or Spearman) for ranking variables. These filters are designed for speed. ttest_filter uses the Rfast package, wilcoxon_filter (Mann-Whitney) test uses matrixTests::row_wilcoxon_twosample, anova_filter uses matrixTests::col_oneway_welch (Welch's F-test) from the matrixTests package. Can be applied to all or a subset of predictors. For mixed datasets (combined continuous & categorical) see stat_filter()

Usage

ttest_filter(
  y,
  x,
  force_vars = NULL,
  nfilter = NULL,
  p_cutoff = 0.05,
  rsq_cutoff = NULL,
  type = c("index", "names", "full"),
  keep_factors = TRUE,
  ...
)

anova_filter(
  y,
  x,
  force_vars = NULL,
  nfilter = NULL,
  p_cutoff = 0.05,
  rsq_cutoff = NULL,
  type = c("index", "names", "full"),
  keep_factors = TRUE,
  ...
)

wilcoxon_filter(
  y,
  x,
  force_vars = NULL,
  nfilter = NULL,
  p_cutoff = 0.05,
  rsq_cutoff = NULL,
  type = c("index", "names", "full"),
  exact = FALSE,
  keep_factors = TRUE,
  ...
)

correl_filter(
  y,
  x,
  method = "pearson",
  force_vars = NULL,
  nfilter = NULL,
  p_cutoff = 0.05,
  rsq_cutoff = NULL,
  type = c("index", "names", "full"),
  keep_factors = TRUE,
  ...
)

Arguments

Value

Integer vector of indices of filtered parameters (type = "index") or character vector of names (type = "names") of filtered parameters in order of t-test p-value. If type is "full" full output from Rfast::ttests is returned.

See Also

lm_filter() stat_filter()

Examples

## sigmoid function
sigmoid <- function(x) {1 / (1 + exp(-x))}

## load iris dataset and simulate a binary outcome
data(iris)
dt <- iris[, 1:4]
colnames(dt) <- c("marker1", "marker2", "marker3", "marker4")
dt <- as.data.frame(apply(dt, 2, scale))
y2 <- sigmoid(0.5 * dt$marker1 + 2 * dt$marker2) > runif(nrow(dt))
y2 <- factor(y2, labels = c("C1", "C2"))

ttest_filter(y2, dt)  # returns index of filtered predictors
ttest_filter(y2, dt, type = "name")  # shows names of predictors
ttest_filter(y2, dt, type = "full")  # full results table

data(iris)
dt <- iris[, 1:4]
y3 <- iris[, 5]
anova_filter(y3, dt)  # returns index of filtered predictors
anova_filter(y3, dt, type = "full")  # shows names of predictors
anova_filter(y3, dt, type = "name")  # full results table

Description

Text progress bar in the R console. Modified from utils::txtProgressBar() to include title and timing.

Usage

txtProgressBar2(
  min = 0,
  max = 1,
  initial = 0,
  char = "=",
  width = NA,
  title = ""
)

Arguments

Details

Use utils::setTxtProgressBar() to set the progress bar and close() to close it.

Value

An object of class "txtProgressBar".

Description

Determines directionality of final predictors for binary or regression models, using the sign of the t-statistic or correlation coefficient respectively for each variable compared to the outcomes.

Usage

var_direction(object)

Arguments

Details

Categorical features with >2 levels are assumed to have a meaningful order for the purposes of directionality. Factors are coerced to ordinal using data.matrix(). If factors are multiclass then directionality results should be ignored.

Value

named vector showing the directionality of final predictors. If the response vector is multinomial NULL is returned.

Description

Uses variable importance across models trained and tested across outer CV folds to assess stability of variable importance. For glmnet, variable importance is measured as the absolute model coefficients, optionally scaled as a percentage. The frequency with which each variable is selected in outer folds as well as the final model is also returned which is helpful for sparse models or with filters to determine how often variables end up in the model in each fold. For glmnet, the direction of effect is taken directly from the sign of model coefficients. For caret models, direction of effect is not readily available, so as a substitute, the directionality of each predictor is determined by the function var_direction() using the sign of a t-test for binary classification or the sign of regression coefficient for continuous outcomes (not available for multiclass caret models). To better understand direction of effect of each predictor within the final model, we recommend using SHAP values - see the vignette "Explaining nestedcv models with Shapley values". See pred_train() for an example.

Usage

var_stability(x, ...)

## S3 method for class 'nestcv.glmnet'
var_stability(
  x,
  ranks = FALSE,
  summary = TRUE,
  percent = TRUE,
  level = 1,
  sort = TRUE,
  ...
)

## S3 method for class 'nestcv.train'
var_stability(x, ranks = FALSE, summary = TRUE, sort = TRUE, ...)

## S3 method for class 'repeatcv'
var_stability(x, ...)

Arguments

Details

Note that for caret models caret::varImp() may require the model package to be fully loaded in order to function. During the fitting process caret often only loads the package by namespace.

Value

If ranks is FALSE and summary is TRUE, returns a dataframe containing mean, sd, sem of variable importance and frequency by which each variable is selected in outer folds. If summary is FALSE, a matrix of either variable importance or, if ranks = TRUE, rankings across the outer folds and the final model is returned, with variables in rows and folds in columns.

See Also

cv_coef() cv_varImp() pred_train()

Description

Calculate weights for class imbalance

Usage

weight(y)

Arguments

Value

Vector of weights