r-s2net

The Generalized Semi-Supervised Elastic-Net.

s2net

Overview

R package s2net

• Our method extends the supervised elastic-net problem, and thus it is a practical solution to the problem of feature selection in semi-supervised contexts.
• Its mathematical formulation is presented from a general perspective, covering a wide range of models.
• We develop a flexible and fast implementation for s2net in R, written in C++ using RcppArmadillo and integrated into R via Rcpp modules.

Installation

You can install the released version of s2net from CRAN with:

install.packages("s2net")

The development version can be installed with:

devtools::install_github("jlaria/s2net", build_vignettes = TRUE)

Features

Example

This is a basic example which shows you how to use the package. Detailed examples can be found in the documentation and vignettes.

library(s2net)
# Auto-MPG dataset is included for benchmark
data("auto_mpg")

Semi-supervised data is made of a labeled dataset xL, the labels yL, and unlabeled data xU. Package s2net includes the function s2Data to process semi-supervised datasets.

head(auto_mpg$P2$xL, 2) # labeled data
#>    displacement horsepower weight acceleration year origin
#> 15          113         91   2372         15.0   70      3
#> 19           97         84   2130         14.5   70      3
head(auto_mpg$P2$yL, 2) # labels
#> [1] 24 27
head(auto_mpg$P2$xU, 2) # unlabeled data
#>   displacement horsepower weight acceleration year origin
#> 1          307         17   3504         12.0   70      1
#> 2          350         35   3693         11.5   70      1

train = s2Data(auto_mpg$P2$xL, auto_mpg$P2$yL, auto_mpg$P2$xU, preprocess = TRUE)

head(train$xL, 2)
#>    displacement horsepower     weight acceleration      year    origin2
#> 15    0.1788500  1.0544632  0.1799762   -0.6392182 -1.878311 -0.6575667
#> 19   -0.5510247  0.7397884 -0.5209896   -0.8471591 -1.878311 -0.6575667
#>     origin3
#> 15 1.356622
#> 19 1.356622

The data is centered and scaled, and factor variables are automatically converted to numerical dummies. Constant columns are also removed. If we wanted to use validation/test data, we must pre-process it according to the training data, with:

valid = s2Data(auto_mpg$P2$xU, auto_mpg$P2$yU, preprocess = train)

There are two ways to fit a semi-supervised elastic-net using s2net. The easiest way is using the function s2netR, that returns an object of S3 class s2netR.

model = s2netR(train, params = s2Params(lambda1 = 0.01, lambda2 = 0.01, gamma1 = 0.01, gamma2 = 100, gamma3 = 0.1))

class(model)
#> [1] "s2netR"
model$beta
#>             [,1]
#> [1,] -0.28152012
#> [2,]  0.04116177
#> [3,] -3.02848437
#> [4,]  0.61602553
#> [5,]  3.65674054
#> [6,]  0.71547766
#> [7,]  0.43169913

ypred = predict(model, valid$xL)

If we are fitting the semi-supervised elastic-net many times, using the same train data (for example, searching for the best hyperparameters), then it is faster to use the C++ class s2net instead.

obj = new(s2net, train, 0) # 0 = linear

# We fit the model with
obj$fit(s2Params(0.01, 0.01, 0.01, 100, 0.1), 0, 2) # frame = 0 (ExtJT), proj = 2 (auto)

obj$beta
#>             [,1]
#> [1,] -0.28700933
#> [2,]  0.04228791
#> [3,] -3.02580178
#> [4,]  0.61559052
#> [5,]  3.65723926
#> [6,]  0.71451133
#> [7,]  0.43040118

ypred = obj$predict(valid$xL, 0) #0=default predictions
# or
ypred = predict(obj, valid$xL)
#> Warning in if (class(newX) == "s2Data") {: the condition has length > 1 and only
#> the first element will be used

Further examples

More examples can be found in the package documentation and vignettes.

vignette(package="s2net")