predict.fk_ppr: Predict method for class fk_ppr

In FKSUM: Fast Kernel Sums

Predict method for class fk_ppr

Description

Standard prediction method for regression models, specific to outputs from the fk_ppr() function. See help(fk_ppr) for more details on the model.

Usage

## S3 method for class 'fk_ppr'
predict(object, Xtest = NULL, ...)

Arguments

object	an object of class fk_ppr, output from the function of the same name.
Xtest	(optional) matrix of test data on which predictions are to be made. If omitted then fitted values from training data are returned.
...	(optional) further arguments passed to or from other methods.

Value

A vector of predictions for Xtest.

Examples

op <- par(no.readonly = TRUE)

set.seed(2)

### Generate a set of data

X = matrix(rnorm(10000), ncol = 10)

### Generate some "true" projection vectors

beta1 = (runif(10)>.5)*rnorm(10)
beta2 = (runif(10)>.5)*rnorm(10)

### Generate responses, dependent on X%*%beta1 and X%*%beta2

y = 1 + X%*%beta1 + ((X%*%beta2)>2)*(X%*%beta2-2)*10
y = y + (X%*%beta1)*(X%*%beta2)/5 + rnorm(1000)

### Fit a PPR model with 2 terms on a sample of the data

train_ids = sample(1:1000, 500)

model = fk_ppr(X[train_ids,], y[train_ids], nterms = 2)

### Predict on left out data, and compute
### estimated coefficient of determination

yhat = predict(model, X[-train_ids,])

MSE = mean((yhat-y[-train_ids])^2)
Var = mean((y[-train_ids]-mean(y[-train_ids]))^2)

1-MSE/Var


#################################################

### Add some "outliers" in the training data and fit
### the model again, as well as one with an absolute loss

y[train_ids] = y[train_ids] + (runif(500)<.05)*(rnorm(500)*100)

model1 <- fk_ppr(X[train_ids,], y[train_ids], nterms = 2)

model2 <- fk_ppr(X[train_ids,], y[train_ids], nterms = 2,
    loss = function(y, hy) abs(y-hy),
    dloss = function(y, hy) sign(hy-y))

### Plot the resulting components in the model on the test data

par(mar = c(2, 2, 2, 2))
par(mfrow = c(2, 2))

plot(X[-train_ids,]%*%model1$vs[1,], y[-train_ids])
plot(X[-train_ids,]%*%model1$vs[2,], y[-train_ids])
plot(X[-train_ids,]%*%model2$vs[1,], y[-train_ids])
plot(X[-train_ids,]%*%model2$vs[2,], y[-train_ids])

par(op)

### estimate comparative estimated coefficients of determination

MSE1 = mean((predict(model1, X[-train_ids,])-y[-train_ids])^2)
MSE2 = mean((predict(model2, X[-train_ids,])-y[-train_ids])^2)
Var = mean((y[-train_ids]-mean(y[-train_ids]))^2)


1-MSE1/Var
1-MSE2/Var

FKSUM documentation built on April 15, 2023, 5:06 p.m.