Note: This package has been maintained by @terrytangyuan since 2015. Please consider sponsoring!

R package for performing and visualizing Local Fisher Discriminant Analysis, Kernel Local Fisher Discriminant Analysis, and Semi-supervised Local Fisher Discriminant Analysis. It's the first package with those methods implemented in native R language. It also provides visualization functions to easily visualize the dimension reduction results.

Introduction to the algorithms and their application can be found here and here. These methods are widely applied in feature extraction, dimensionality reduction, clustering, classification, information retrieval, and computer vision problems. An introduction to the package is also available in Chinese here.

Welcome any feedback and pull request.

install.packages('lfda')

devtools::install_github('terrytangyuan/lfda')

Please call citation("lfda") in R to properly cite this software. A white paper is published on Journal of Open Source Software here.

Suppose we want to reduce the dimensionality of the original data set (we are using iris data set here) to 3, then we can run the following:

k <- iris[,-5] # this matrix contains all the predictors to be transformed
y <- iris[,5] # this should be a vector that represents different classes
r <- 3 # dimensionality of the resulting matrix

# run the model, note that two other kinds metrics we can use: 'weighted' and 'orthonormalized'
model <- lfda(k, y, r, metric = "plain") 

plot(model, y) # 3D visualization of the resulting transformed data set

predict(model, iris[,-5]) # transform new data set using predict

The main usage is the same except for an additional kmatrixGauss call to the original data set to perform a kernel trick:

k <- kmatrixGauss(iris[,-5])
y <- iris[,5]
r <- 3
model <- klfda(k, y, r, metric = "plain")

Note that the predict method for klfda is still under development. The plot method works the same way as in lfda.

This algorithm requires one additional argument such as beta that represents the degree of semi-supervisedness. Let's assume we ignore 10% of the labels in iris data set:

k <- iris[,-5]
y <- iris[,5]
r <- 3
model <- self(k, y, beta = 0.1, r = 3, metric = "plain")

The methods predict and plot work the same way as in lfda.

{ggplot2::autoplot} has been integrated with this package. Now {lfda} can be plotted in 2D easily and beautifully using {ggfortify} package. Go to this link and scroll down to the last section for an example.

To contribute to this project, please take a look at the Contributing Guidelines first. Please note that this project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.

Contact the maintainer of this package: Yuan Tang [email protected]