ISODATA.md

ISODATA Algorithm CIsoDataClustering

• ISODATA Algorithm CIsoDataClustering
  • Parameters
  • Sample

This method is a heuristic algorithm based on geometrical proximity of the data points. For the description, see Ball, Geoffrey H., Hall, David J. Isodata: a method of data analysis and pattern classification. (1965)

The final result depends greatly on the initial settings.

In NeoML the algorithm is implemented by the CIsoDataClustering class that provides the IClustering interface. Its Clusterize method is used to split the data set into clusters.

Parameters

The clustering parameters are described by the CIsoDataClustering::CParam structure.

• InitialClustersCount — the initial number of clusters
• MaxClustersCount — the maximum number of clusters that can be reached during processing
• MinClusterSize — the minimum cluster size
• MaxIterations — the maximum number of iterations
• MinClustersDistance — the minimum distance between the clusters (if it is smaller, the clusters may be merged)
• MaxClusterDiameter — the maximum cluster diameter (if it is exceeded the cluster may be split)
• MeanDiameterCoef — how much the cluster diameter may exceed the mean diameter across all the clusters (if a cluster diameter is larger than the mean diameter multiplied by this value it may be split)

Sample

This sample shows how to use the ISODATA algorithm to clusterize the Iris Data Set:

void Clusterize( IClusteringData& irisDataSet, CClusteringResult& result )
{
	CIsoDataClustering::CParam params;
	params.InitialClustersCount = 1;
	params.MaxClustersCount = 20;
	params.MinClusterSize = 1;
	params.MinClustersDistance = 0.60;
	params.MaxClusterDiameter = 1.0;
	params.MeanDiameterCoef = 0.5;
	params.MaxIterations = 50;

	CIsoDataClustering isoData( params );
	isoData.Clusterize( irisDataSet, result );
}