kitchen_prediction: Make predictions with Convolutional Kitchen Sinks

In avery-kruger/kitchen: Convolutional Kitchen Sinks in R

Make predictions with Convolutional Kitchen Sinks

Description

'kitchen_prediction' trains models on CKS-projected training data and then predicts values for some other data using those models.

Usage

kitchen_prediction(
  trainx,
  trainy,
  predictx,
  features,
  windows,
  verbose = TRUE,
  simplify = FALSE,
  clampoutliers = TRUE,
  bootstrap = NULL,
  seed = NULL,
  write_progress = NULL,
  ...
)

Arguments

trainx	A matrix of independent data to use for training models. trainx[i,] should correspond with trainy[i].
trainy	A vector of dependent data to use for training models. trainy[i] should correspond with trainx[i,].
predictx	A vector or matrix to predict values for. Values must correspond with the training data.
features	A value or vector of feature counts to use in the kitchen sink models.
windows	A value or vector of window sizes to use in the kitchen sink models. Values should not exceed the column length of the data.
verbose	Print progress.
simplify	Return a matrix rather than a list when there is only one combination of feature count and window size.
clampoutliers	Clamp predicted values beyond range of training values to the minimum or maximum of the training values.
bootstrap	A number of times to bootstrap predictions by predicting from data sampled with replacement.
seed	Set a seed for repeatable norms.
write_progress	A file name to write predictions to.
...	Arguments to be passed to kitchen_sink(). Can include a bias term or alternate nonlinearization functions.

Details

This function has a few uses. Most simply, it can generate predictions for data from a single kitchen sink model when provided with a single set of hyperparameters (feature count and window size). kitchen_sweep() can quickly assess what hyperparameters perform best. Models are trained by ridge regression using cv.glmnet, allowing for reduction of overfitting. When 'reps' is greater than one, this function makes multiple predictions using unique normal matrices; the set of predictions can be used for confidence intervals of the true kernel function. If a set of hyperparameters are provided, the function will generate predictions for each unique combination, which could allow for model averaging.

Value

Returns a list where index [[f]][[w]] returns the predicted values from a model with feature count features[f] and window size windows[w].

Author(s)

Avery Kruger

See Also

kitchen_sweep()

Examples

x <- matrix(sample(1:10,10000,TRUE),2000,5)
y <- 5*x[,1] + 20*x[,1]*x[,2] + 3*x[,3]^2 - 10*x[,4] - 2*x[,5]
kitchen_sweep(x[1:1000,],y[1:1000],
   x[1001:2000,],y[1001:2000],
   2^(4:8),2:5)

a <- matrix(sample(1:10,1000,TRUE),1,5)
b <- 5*a[,1] + 20*a[,1]*a[,2] + 3*a[,3]^2 - 10*a[,4] - 2*a[,5]
mybootstrap <- kitchen_prediction(x,y,a,64,5,bootstrap=10)
hist(unlist(mybootstrap))

avery-kruger/kitchen documentation built on May 7, 2024, 6:40 a.m.