mlNnet {mlearning}

Supervised classification and regression using neural network

Description

Unified (formula-based) interface version of the single-hidden-layer neural network algorithm, possibly with skip-layer connections provided by nnet::nnet().

Usage

mlNnet(train, ...)

ml_nnet(train, ...)

## S3 method for class 'formula'
mlNnet(
  formula,
  data,
  size = NULL,
  rang = NULL,
  decay = 0,
  maxit = 1000,
  ...,
  subset,
  na.action
)

## Default S3 method:
mlNnet(train, response, size = NULL, rang = NULL, decay = 0, maxit = 1000, ...)

## S3 method for class 'mlNnet'
predict(
  object,
  newdata,
  type = c("class", "membership", "both", "raw"),
  method = c("direct", "cv"),
  na.action = na.exclude,
  ...
)

Arguments

train

a matrix or data frame with predictors.

...

further arguments passed to nnet::nnet() that has many more parameters (see its help page).

formula

a formula with left term being the factor variable to predict (for supervised classification), a vector of numbers (for regression) and the right term with the list of independent, predictive variables, separated with a plus sign. If the data frame provided contains only the dependent and independent variables, one can use the class ~ . short version (that one is strongly encouraged). Variables with minus sign are eliminated. Calculations on variables are possible according to usual formula convention (possibly protected by using I()).

data

a data.frame to use as a training set.

size

number of units in the hidden layer. Can be zero if there are skip-layer units. If NULL (the default), a reasonable value is computed.

rang

initial random weights on [-rang, rang]. Value about 0.5 unless the inputs are large, in which case it should be chosen so that rang * max(|x|) is about 1. If NULL, a reasonable default is computed.

decay

parameter for weight decay. Default to 0.

maxit

maximum number of iterations. Default 1000 (it is 100 in nnet::nnet()).

subset

index vector with the cases to define the training set in use (this argument must be named, if provided).

na.action

function to specify the action to be taken if NAs are found. For ml_nnet() na.fail is used by default. The calculation is stopped if there is any NA in the data. Another option is na.omit, where cases with missing values on any required variable are dropped (this argument must be named, if provided). For the predict() method, the default, and most suitable option, is na.exclude. In that case, rows with NAs in ⁠newdata=⁠ are excluded from prediction, but reinjected in the final results so that the number of items is still the same (and in the same order as ⁠newdata=⁠).

response

a vector of factor (classification) or numeric (regression).

object

an mlNnet object

newdata

a new dataset with same conformation as the training set (same variables, except may by the class for classification or dependent variable for regression). Usually a test set, or a new dataset to be predicted.

type

the type of prediction to return. "class" by default, the predicted classes. Other options are "membership" the membership (number between 0 and 1) to the different classes, or "both" to return classes and memberships. Also type "raw" as non normalized result as returned by nnet::nnet() (useful for regression, see examples).

method

"direct" (default) or "cv". "direct" predicts new cases in ⁠newdata=⁠ if this argument is provided, or the cases in the training set if not. Take care that not providing ⁠newdata=⁠ means that you just calculate the self-consistency of the classifier but cannot use the metrics derived from these results for the assessment of its performances. Either use a different data set in ⁠newdata=⁠ or use the alternate cross-validation ("cv") technique. If you specify method = "cv" then cvpredict() is used and you cannot provide ⁠newdata=⁠ in that case.

Value

ml_nnet()/mlNnet() creates an mlNnet, mlearning object containing the classifier and a lot of additional metadata used by the functions and methods you can apply to it like predict() or cvpredict(). In case you want to program new functions or extract specific components, inspect the "unclassed" object using unclass().

See Also

mlearning(), cvpredict(), confusion(), also nnet::nnet() that actually does the classification.

Examples

# Prepare data: split into training set (2/3) and test set (1/3)
data("iris", package = "datasets")
train <- c(1:34, 51:83, 101:133)
iris_train <- iris[train, ]
iris_test <- iris[-train, ]
# One case with missing data in train set, and another case in test set
iris_train[1, 1] <- NA
iris_test[25, 2] <- NA

set.seed(689) # Useful for reproductibility, use a different value each time!
iris_nnet <- ml_nnet(data = iris_train, Species ~ .)
summary(iris_nnet)
predict(iris_nnet) # Default type is class
predict(iris_nnet, type = "membership")
predict(iris_nnet, type = "both")
# Self-consistency, do not use for assessing classifier performances!
confusion(iris_nnet)
# Use an independent test set instead
confusion(predict(iris_nnet, newdata = iris_test), iris_test$Species)

# Idem, but two classes prediction
data("HouseVotes84", package = "mlbench")
set.seed(325)
house_nnet <- ml_nnet(data = HouseVotes84, Class ~ ., na.action = na.omit)
summary(house_nnet)
# Cross-validated confusion matrix
confusion(cvpredict(house_nnet), na.omit(HouseVotes84)$Class)

# Regression
data(airquality, package = "datasets")
set.seed(74)
ozone_nnet <- ml_nnet(data = airquality, Ozone ~ ., na.action = na.omit,
  skip = TRUE, decay = 1e-3, size = 20, linout = TRUE)
summary(ozone_nnet)
plot(na.omit(airquality)$Ozone, predict(ozone_nnet, type = "raw"))
abline(a = 0, b = 1)