nnetModel: Combine Model-based Recursive Partitioning with Neural...
In schiffner/locClass: Collection of Local Classification Methods

Description Usage Arguments Format Details Value References See Also Examples

Combine Model-based Recursive Partitioning with Neural Networks.

nnetModel

\method{reweight}{nnetModel}(object, weights, ...)

\method{deviance}{nnet}(object, ...)

\method{estfun}{nnet}(x, ...)

predict.nnetModel(object, out = c("class", "posterior"), newdata, ...)

`object`	An object of class "nnetModel" and "nnet", respectively.
`weights`	A vector of observation weights.
`x`	An object of class "nnet".
`out`	Should class labels or posterior probabilities be returned?
`newdata`	A `data.frame` of cases to be classified.
`...`	Further arguments (e.g. to `nnet` and `nnetRep`).

An object of class StatModel of length 1.

This page lists all ingredients to combine Neural Networks with Model-Based Recursive Partitioning (mob from package party). See the example for how to do that.

nnetModel is an object of class StatModel-class implemented in package modeltools that provides an infra-structure for an unfitted nnet model.

Moreover, methods for nnet and nnetModel objects for the generic functions reweight, deviance, estfun, and predict are provided.

reweight: The re-weighted fitted "nnetModel" object.
deviance: The value of the objective function extracted from object.
estfun: The empirical estimating (or score) function, i.e. the derivatives of the objective function with respect to the parameters, evaluated at the training data.
predict: Either a vector of predicted class labels or a matrix of class posterior probabilities.

Zeileis, A., Hothorn, T. and Kornik, K. (2008), Model-based recursive partitioning. Journal of Computational and Graphical Statistics, 17(2) 492–514.

reweight, deviance, estfun, predict.

library(benchData)

data <- vData(500)
x <- seq(0,1,0.05)
grid <- expand.grid(x.1 = x, x.2 = x)

fit <- mob(y ~ x.1 + x.2 | x.1 + x.2, data = data, model = nnetModel, size = 1, trace = FALSE,
control = mob_control(objfun = deviance, minsplit = 200))

## predict posterior probabilities
pred <- predict(fit, newdata = grid, out = "posterior")
post <- do.call("rbind", pred)

image(x, x, matrix(as.numeric(post[,1]), length(x)), xlab = "x.1", ylab = "x.2")
contour(x, x, matrix(as.numeric(post[,1]), length(x)), levels = 0.5, add = TRUE)
points(data$x, pch = as.character(data$y))

## predict node membership
splits <- predict(fit, newdata = grid, type = "node")
contour(x, x, matrix(splits, length(x)), levels = min(splits):max(splits), add = TRUE, lty = 2)

## training error
mean(predict(fit) != data$y)