cv_sgdnet: _k_-fold Cross Validation for sgdnet
In jolars/sgdnet: Fast Sparse Linear Models for Big Data with SAGA

Description Usage Arguments Details Value Measures Author(s) See Also Examples

This function performs model validation by k-fold cross validation for models fit with sgdnet() over the entire regularization path and/or various elastic net penalties.

1
2
3

cv_sgdnet(x, y, alpha = 1, lambda = NULL, nfolds = 10,
  foldid = NULL, type.measure = c("deviance", "mse", "mae", "class",
  "auc"), ...)

`x`	input matrix
`y`	response variable
`alpha`	elastic net mixing parameter; vectors of values are allowed (unlike in `sgdnet()`)
`lambda`	regularization strength
`nfolds`	number of folds (k) – 3 is the minimum allowed
`foldid`	a vector of fold identities of the same length as the number of observations
`type.measure`	the type of error, one of ("deviance" (default), "mse" (mean-squared error), "mae" (mean absolute error), "class" (misclassification rate), "auc" (area under the curve)); see Measures to find out what types each family has available.
`...`	arguments passed on to `sgdnet()`

The primary usage of this model is to tune for values of lambda and alpha. This function will randomly divide the data into k folds. For each fold, the remaining k-1 will be used to train a model across a regularization path, and optionally a range of alpha. The fold that is left out is then used to measure the performance of the model. We proceed across all the folds, which means that each observation is used exactly once for validation, and finally average our results across all the folds.

An object of class 'cv_sgdnet' with the following items:

`alpha`	the elastic net mixing parameter used
`lambda`	a list of lambda values of the same length as `alpha`
`cv_summary`	a `data.frame` summarizing the prediction error across the regularization path with columns `alpha`, `lambda`, `mean`, `sd`, `ci_lo`, `ci_up`
`cv_raw`	the raw cross-validation scores as a list of the same length as `alpha`, each item a `matrix` with the error for each fold as a row and each value of `lambda` in columns.
`name`	the type of prediction error used
`fit`	a fit from `sgdnet()` to the full data set based on the `alpha` with the best cross-validation score
`alpha_min`	the `alpha` corresponding to the fit with the best cross-validation performance
`lambda_min`	the `lambda` corresponding to the fit with the best cross-validation performance
`lambda_1se`	the largest `lambda` with a cross-validation performance within one standard deviation of the one coresponding to `lambda_min`

Family	`deviance`	`mse`	`mae`	`class`	`auc`
`gaussian`	x (`mse`)	x	x
`binomial`	x	x	x	x	x
`multinomial`	x	x	x	x
`mgaussian`	x (`mse`)	x	x

Johan Larsson (partly consisting of modified code from glmnet::cv.glmnet() by Jerome Friedman, Trevor Hastie, Rob Tibshirani, and Noah Simon)

sgdnet(), predict.cv_sgdnet(), plot.cv_sgdnet()

set.seed(1)
n <- nrow(heart$x)
train_ind <- sample(n, floor(0.8*n))
cv_fit <- cv_sgdnet(heart$x[train_ind, ],
                    heart$y[train_ind],
                    family = "binomial",
                    nfolds = 7,
                    alpha = c(0, 1))
plot(cv_fit)
predict(cv_fit, heart$x[-train_ind, ], s = "lambda_min")