Implementations of Semi-Supervised Learning Approaches for Classification

A collection of implementations of semi-supervised classifiers and methods to evaluate their performance. The package includes implementations of, among others, Implicitly Constrained Learning, Moment Constrained Learning, the Transductive SVM, Manifold regularization, Maximum Contrastive Pessimistic Likelihood estimation, S4VM and WellSVM.

This R package provides implementations of several semi-supervised learning methods, in particular, our own work involving constraint based semi-supervised learning.

The package is still under development. Therefore, function names and interfaces are subject to change.

To cite the package, use either of these two references:

• Krijthe, J.H. & Loog, M. (2015). Implicitly Constrained Semi-Supervised Least Squares Classification. In E. Fromont, T. de Bie, & M. van Leeuwen, eds. 14th International Symposium on Advances in Intelligent Data Analysis XIV (Lecture Notes in Computer Science Volume 9385). Saint Etienne. France, pp. 158-169.
• Jesse H. Krijthe (2016). RSSL: Implementations of Semi-Supervised Learning Approaches for Classification, URL: https://github.com/jkrijthe/RSSL

Installation Instructions

This package is not yet available on CRAN. To install the latest version of the package using the devtools package:

library(devtools)
install_github("jkrijthe/RSSL")

Usage

After installation, load the package as usual:

library(RSSL)

The following code generates a simple dataset, trains a supervised and two semi-supervised classifiers and evaluates their performance:

library(dplyr,warn.conflicts = FALSE)
library(ggplot2,warn.conflicts = FALSE)
 
set.seed(2)
df <- generate2ClassGaussian(200, d=2, var = 0.2, expected=TRUE)
df <- df %>% add_missinglabels_mar(Class~.,prob=0.98) 
 
# Train classifier
g_nm <- NearestMeanClassifier(Class~.,df,prior=matrix(0.5,2))
g_self <- SelfLearning(Class~.,df,
                       method=NearestMeanClassifier,
                       prior=matrix(0.5,2))
 
# Plot dataset
df %>% 
  ggplot(aes(x=X1,y=X2,color=Class,size=Class)) +
  geom_point() +
  coord_equal() +
  scale_size_manual(values=c("-1"=3,"1"=3), na.value=1) +
  geom_linearclassifier("Supervised"=g_nm,
                  "Semi-supervised"=g_self)

 
# Evaluate performance: Squared Loss & Error Rate
mean(loss(g_nm,df))
mean(loss(g_self,df))
 
 
mean(predict(g_nm,df)!=df$Class)
mean(predict(g_self,df)!=df$Class)

Acknowledgement

Work on this package was supported by Project 23 of the Dutch national program COMMIT.