Enhanced Mode Clustering

Performing mode clustering and visualization.

• Paper reference: Chen, Yen-Chi, Christopher R. Genovese, and Larry Wasserman. "A comprehensive approach to mode clustering." Electronic Journal of Statistics 10.1 (2016): 210-241.
• Maintianer: yenchic@uw.edu
• EMC.R: the main script for running the program.
• Example_OliveOil.R: an example for applying EMC method to the Olive Oil dataset.

EMC.R

Here is a description for the main functions.

fms

fms = function(data, query, h, eps=1.0e-8, max.iterations=100, cut = 0.1)

• Fast mean shift using heirachical clustering.
• Inputs:
  • data: Input data matrix.
  • query: The mesh points that you want to apply mean shift to.
  • h: Smoothing parameter.
  • max.iterations: Maximal number of iteration for mean shift.
  • eps: The tolerance. If mean shift moves less than this value, we will consider it done.
  • cut: The cut for heirachical clustering (we cut the dedrogram by height = cut*h).
• Outputs:
  • The mode clustering result; a list consisting of
    • label: The cluster labels for query points.
    • modes: The local modes corresponding to each label.

EMC

EMC = function(data, h=NULL, eps=1.0e-8, max.iterations=100, ...) UseMethod("EMC")

EMC.default = function(data, h=NULL, eps=1.0e-8, max.iterations=100, n0= NULL, rho=NULL, cut=0.1, noisy = F, T_denoise =5)

• Enhanced mode clustering.
• Inputs:
  • data: Input data matrix.
  • h: Smoothing parameter.
  • max.iterations: Maximal number of iteration for mean shift.
  • eps: The tolerance. If mean shift moves less than this value, we will consider it done.
  • n0: The thresholding size for tiny clusters. Default is to use the method given in Chen et al. (2014).
  • rho: The contrast parameter for visualization. Default is to use the method given in Chen et al. (2014).
  • noisy: True or False. To desplay noisy clusters (without thresholding). Default is False.
  • T_denoise: Maximal number of denoising. If tiny clusters presence, we will remove them and redo mode clustering. This is the maximal number of redoing mean shift clustering.
  • cut: The cut for heirachical clustering (we cut the dedrogram by height = cut*h).
• Output:
  • An S4 object about summary informations using enhanced mode clustering. A list consisting:
    • label: The cluster labels for query points.
    • modes: The local modes corresponding to each label.
    • c.matrix: The connectivity matrix.
    • vis.data: The visualization coordinates for data points.
    • vis.modes: The visualization coordinates for local modes.
    • SC.plot: The size of ordered clusters before denoising.
    • size.threshold: The size threshold for denoising tiny clusters.
    • bandwidth: The smoothing bandwidth.
    • rho: The contrast paramter used for visualization.
    • noisy.label: The cluster labels for query points before denoising.
    • noisy.modes: The local modes corresponding to each label before denoising.