"NeighborhoodContraction" (Machine Learning Method)

Details & Suboptions

  • "NeighborhoodContraction" is a neighbor-based clustering method. "NeighborhoodContraction" works for arbitrary cluster shapes and sizes, however, it can fail when clusters have different densities or are intertwined.
  • The following plots show the results of the "NeighborhoodContraction" method applied to toy datasets:
  • The "NeighborhoodContraction" method iteratively shifts data points toward higher density regions. During this procedure, data points tend to collapse to different fixed points, each of them representing a cluster.
  • Formally, at each step, each data point is set to the mean of its neighboring points , .
  • Neighboring points are defined as all the points within a ball of ϵ radius. The algorithm repeats the updates until points stop moving; all points belonging to a cluster are then collapsed (up to a tolerance). This algorithm is equivalent to the "MeanShift" method but with a different neighborhood definition.
  • The following suboption can be given:
  • "NeighborhoodRadius"Automaticradius ϵ

Examples

open allclose all

Basic Examples  (3)

Find clusters of nearby values using the "NeighborhoodContraction" clustering method:

Train the ClassifierFunction on a list of colors using the "NeighborhoodContraction" method:

Gather the elements by their class number:

Train a ClassifierFunction on a mixture of two-dimensional normal distributions:

Find the cluster assignments and visualize them:

Options  (3)

DistanceFunction  (2)

Generate a list of 200 random colors:

Find clusters by specifying a DistanceFunction option:

Generate points based on a mixture of two-dimensional normal distributions:

Find different clustering structures by specifying different DistanceFunction options:

"NeighborhoodRadius"  (1)

Generate points based on a mixture of two-dimensional normal distributions:

Find different clustering structures by specifying different "NeighborhoodRadius" suboptions:

Introduced in 2020
 (12.1)