U-Rerf

The unsupervised version of Rerf. Currently U-Rerf creates either an NxN similarity matrix which shows the similarity between each of the N samples to each of the other N-1 samples or U-Rerf can create an NxN distance matrix showing the distance between each of the N samples to each of the other N samples.

U-Rerf Setup

• Clone the repository
• source the U-Rerf functions -- source('rfr_us.R')
• To run the examples
  • install ggplot2 -- install.packages("ggplot2")
  • install scatterplot3d -- install.packages("scatterplot3d")

Usage

To create a similarity matrix use the createSimilarityMatrix function which is loaded when the rfr_us.R file is sourced. The function takes three inputs:

1. X - which is an N row by d column matrix of N samples each of d dimensions. X must be numeric.
2. numTrees - the number of trees used to create the forest. 100 is the default.
3. K - this is really the minparent used in creating the forest but should be thought of as the K in K nearest neighbors.

Examples

U-Rerf can either create a distance matrix or a similarity matrix. This is a simple example using the Iris dataset which can be copied and pasted into an R terminal-

# make sure to put the correct path to rfr_us.R
source('rfr_us.R')

# number of trees for forest
numtrees <- 100
# the 'k' of k nearest neighbors
k <- 10

# create a sizeD by m synthetic dataset
X <- as.matrix(iris[,1:4])

# create a similarity and distance matrices using urerf
similarityMatrix <- createSimilarityMatrix (X, numtrees, k)
distanceMatrix <- createDistanceMatrix (X, numtrees, k)

In addition, several examples are included with corresponding outputs-

1. 1 dimensional line. 1dimensionalNN.R -- 1dimDVN.pdf
2. 3 dimensional line. 3dimensionalNN.R -- 3dimDvN.pdf
3. 10 dimensional line. 10dimensionalNN.R -- 10dimDvN.pdf
4. 3 dimensional line with an outlier. 3dimensionalNNwOutlier.R -- 3dimDvNwO.pdf
5. Swiss Roll. 3dSwissRoll.R -- 3dswissDvN.pdf
6. Swiss Roll with an outlier. 3dSwissRollwOutlier.R -- 3dimDvNwO.pdf