| Title: | Real-Time Adaptive Penalization for Streaming Lasso Models |
|---|---|
| Description: | An implementation of the Real-time Adaptive Penalization (RAP) algorithm through which to iteratively update a regularization parameter in a streaming context. |
| Authors: | Ricardo Pio Monti |
| Maintainer: | Ricardo Pio Monti <[email protected]> |
| License: | GPL-2 |
| Version: | 1.1 |
| Built: | 2024-11-14 06:14:39 UTC |
| Source: | https://github.com/cran/rRAP |
This package provides an implementation of the Real-time adaptive penalization (RAP) algorithm through which to iteratively update a regularization parameter in a streaming context.
| Package: | rRAP |
| Type: | Package |
| Version: | 1.0 |
| Date: | 2016-09-29 |
| License: | GPL-2 |
Ricardo Pio Monti Maintainer: Ricardo Pio Monti <[email protected]>
See Monti et al, "A framework for adaptive regularization in streaming Lasso models", 2016
# Recreate Figure 1 from Monti et al 2016 library(lars) data(diabetes) Data = cbind(diabetes$y, diabetes$x) # initialize RAP object R = RAP(X = matrix(diabetes$x[1,], nrow=1), y = diabetes$y[1], r = .995, eps = 0.0005, l0 = .1) # iteratively update: ## Not run: for (i in 2:nrow(Data)){ R = update.RAP(RAPobj=R, Ynew = diabetes$y[i], Xnew=matrix(diabetes$x[i,], nrow=1)) } ## End(Not run)# Recreate Figure 1 from Monti et al 2016 library(lars) data(diabetes) Data = cbind(diabetes$y, diabetes$x) # initialize RAP object R = RAP(X = matrix(diabetes$x[1,], nrow=1), y = diabetes$y[1], r = .995, eps = 0.0005, l0 = .1) # iteratively update: ## Not run: for (i in 2:nrow(Data)){ R = update.RAP(RAPobj=R, Ynew = diabetes$y[i], Xnew=matrix(diabetes$x[i,], nrow=1)) } ## End(Not run)
Obtain prediction based on current estimate of sparse linear regression model
## S3 method for class 'RAP' predict(object, Xnew, ...)## S3 method for class 'RAP' predict(object, Xnew, ...)
object |
Current RAP object |
Xnew |
New observation from which to predict |
... |
Additional arguments |
Produces a matrix of predicted values
Ricardo Pio Monti
Monti et al, "A framework for adaptive regularization in streaming Lasso models", 2016
# library(lars) data(diabetes) Data = cbind(diabetes$y, diabetes$x) # initialize RAP object with a burn in of 50 observations R = RAP(X = matrix(diabetes$x[1:50,], nrow=50), y = diabetes$y[1:50], r = .995, eps = 0.0005, l0 = .1) # make predictions: #predict.RAP(object = R, Xnew = diabetes$x[50:70,])# library(lars) data(diabetes) Data = cbind(diabetes$y, diabetes$x) # initialize RAP object with a burn in of 50 observations R = RAP(X = matrix(diabetes$x[1:50,], nrow=50), y = diabetes$y[1:50], r = .995, eps = 0.0005, l0 = .1) # make predictions: #predict.RAP(object = R, Xnew = diabetes$x[50:70,])
This function initializes and RAP object. This contains a Lasso regression model together with methods to iteratively update the regularization parameter.
RAP(X, y, r = 0.95, eps = 0.01, l0 = 0.1, Approx = FALSE)RAP(X, y, r = 0.95, eps = 0.01, l0 = 0.1, Approx = FALSE)
X |
Burn in training data. Can either be a single observation (in this case a matrix with 1 row) or several. This must be a matrix. |
y |
Burn in response data |
r |
Fixed forgetting factor used to update |
eps |
Fixed stepsize used to update regularization parameter |
l0 |
Initial guess for regularization parameter |
Approx |
Boolean indicating whether exact or approximate gradient should be calculated when updating regularization parameter. |
See Monti et al, "A framework for adaptive regularization in streaming Lasso models", 2016
A RAP object is returned with the following elements:
r |
Fixed forgetting factor |
eps |
Stepsize used to update regularization parameter |
w |
Current measure of effective sample size |
xbar |
|
St |
|
regParam |
Current estimate of regularization parameter |
l1Track |
Vector storing all past estimates of regularization parameter |
beta |
Current estimate of regression coefficients |
Approx |
Boolean indicating if exact or approximate gradients where employed |
The object has the following methods:
update |
Update regularization parameters and regression coefficients based on new data |
predict |
Predict based on current model |
Warning that this implementation uses the shooting algorithm (co-ordinate gradient descent) to update regression coefficients. A more efficient implementation would employ stochastic gradient descent.
Ricardo Pio Monti
Monti et al, "A framework for adaptive regularization in streaming Lasso models", 2016
# Recreate Figure 1 from library(lars) data(diabetes) Data = cbind(diabetes$y, diabetes$x) # initialize RAP object R = RAP(X = matrix(diabetes$x[1,], nrow=1), y = diabetes$y[1], r = .995, eps = 0.0005, l0 = .1) # iteratively update: ## Not run: for (i in 2:nrow(Data)){ R = update.RAP(RAPobj=R, Ynew = diabetes$y[i], Xnew=matrix(diabetes$x[i,], nrow=1)) } ## End(Not run)# Recreate Figure 1 from library(lars) data(diabetes) Data = cbind(diabetes$y, diabetes$x) # initialize RAP object R = RAP(X = matrix(diabetes$x[1,], nrow=1), y = diabetes$y[1], r = .995, eps = 0.0005, l0 = .1) # iteratively update: ## Not run: for (i in 2:nrow(Data)){ R = update.RAP(RAPobj=R, Ynew = diabetes$y[i], Xnew=matrix(diabetes$x[i,], nrow=1)) } ## End(Not run)
Update regularization parameter and the associated Lasso regression coefficients, Updates can either be mini-batch or single observations.
## S3 method for class 'RAP' update(object, Ynew, Xnew, ...)## S3 method for class 'RAP' update(object, Ynew, Xnew, ...)
object |
Current RAP object |
Ynew |
New response. In the case of mini-batch updates a vector should be provided. |
Xnew |
New covariates. This should be a matrix. |
... |
Additional arguments |
See Monti et al 2016
A RAP objecti is returned where the regularization parameter and the estimated regression coefficients have been updated.
Warning that this implementation uses the shooting algorithm (co-ordinate gradient descent) to update regression coefficients. A more efficient implementation would employ stochastic gradient descent.
Ricardo Pio Monti
See Monti et al, "A framework for adaptive regularization in streaming Lasso models", 2016
# Recreate Figure 1 from library(lars) data(diabetes) Data = cbind(diabetes$y, diabetes$x) # initialize RAP object R = RAP(X = matrix(diabetes$x[1,], nrow=1), y = diabetes$y[1], r = .995, eps = 0.0005, l0 = .1) # iteratively update: ## Not run: for (i in 2:nrow(Data)){ R = update.RAP(object=R, Ynew = diabetes$y[i], Xnew=matrix(diabetes$x[i,], nrow=1)) } ## End(Not run)# Recreate Figure 1 from library(lars) data(diabetes) Data = cbind(diabetes$y, diabetes$x) # initialize RAP object R = RAP(X = matrix(diabetes$x[1,], nrow=1), y = diabetes$y[1], r = .995, eps = 0.0005, l0 = .1) # iteratively update: ## Not run: for (i in 2:nrow(Data)){ R = update.RAP(object=R, Ynew = diabetes$y[i], Xnew=matrix(diabetes$x[i,], nrow=1)) } ## End(Not run)