MapperMDS

Adaptation from TDA.jl mapper tool. Here the input is a distance matrix and it uses a hierarchical clustering. It was inspired by the problems when one only has the distance matrix of a dataset and not the point cloud itself.

Installation

In the julia REPL

using Pkg
Pkg.add("https://github.com/chronchi/MapperMDS.jl.git")

Usage

The usage here is pretty similar to wildart's module. We have to pass a distance matrix and filter values of the distance matrix, e.g., MDS using the distance matrix and projecting to the real line.

The package performs a hierarchical clustering, since it only needs a distance matrix (or a dissimilarity matrix).

The linkage criteria can be given as a parameter to the function mapper via linkage = :criteria, where criteria is one of the options here.

You can choose the method to evaluate the quality of the clustering. The default is no method. You can choose the silhouette using clustselection = MapperMDS.silhouette. You can then specify the number of clusters to the hierarchical clustering with n_clusters = k. The default value is 2. Notice we can use DBScan from the package Clustering. When using DBScan it is important to use the default value of clustselection.

To plot we use the package GraphPlot together with LightGraphs.

Example

using MapperMDS
import LinearAlgebra: rand, norm
import MultivariateStats: classical_mds

# generate a random symmetric matrix with non negative entries
X = rand(100, 100)
X -= X'
X = map(abs, X)

# embedding of the data represented by the distance matrix in R.
filter = classical_mds(X, 1)'[:,1]

# call mapper. note we don't have to specify the filter function,
# as filter is the filtered data already.
mpr = MapperMDS.mapper(X, filter, intervals=5, overlap=0.2)

# plot the corresponding graph
plot_graphmpr(mpr)