org.ddogleg.clustering.kmeans

Class AssignKMeans_F64

java.lang.Object

org.ddogleg.clustering.kmeans.AssignKMeans_F64

All Implemented Interfaces:

java.io.Serializable, AssignCluster<double[]>

public class AssignKMeans_F64
extends java.lang.Object
implements AssignCluster<double[]>

Implementation of AssignCluster for K-Means. Euclidean distance squared is used to select the best fit clusters to a point. This distance metric works well for hard assignment but can produce undesirable results for soft assignment, see JavaDoc.

See Also:

Serialized Form

Constructor Summary

Constructors

Constructor and Description
AssignKMeans_F64(AssignKMeans_F64 original)
AssignKMeans_F64(java.util.List<double[]> clusters)

Method Summary

Modifier and Type	Method and Description
int	assign(double[] point) Assigns the point to cluster which is the best fit.
void	assign(double[] point, double[] fit) Soft assignment is done by summing the total distance of the point from each cluster.
AssignCluster<double[]>	copy() Creates an exact copy of this class.
java.util.List<double[]>	getClusters()
int	getNumberOfClusters() Total number of clusters.
void	setClusters(java.util.List<double[]> clusters)

Methods inherited from class java.lang.Object

equals, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

AssignKMeans_F64

public AssignKMeans_F64(java.util.List<double[]> clusters)

AssignKMeans_F64

public AssignKMeans_F64(AssignKMeans_F64 original)

Method Detail

assign

public int assign(double[] point)

Description copied from interface: AssignCluster

Assigns the point to cluster which is the best fit.

Specified by:

assign in interface AssignCluster<double[]>

Parameters:

point - Point which is to be assigned

Returns:

Index of the cluster from 0 to N-1

assign

public void assign(double[] point,
                   double[] fit)

Soft assignment is done by summing the total distance of the point from each cluster. Then for each cluster its value is set to total minus its distance. The output array is then normalized by dividing each element by the sum.

When all clusters are approximately the same distance or one is clearly the closest this produces reasonable results. When multiple clusters are much closer than at least on other cluster then it effectively ignores the relative difference in distances between the closest points. There are several obvious heuristic "fixes" to this issue, but the best way to solve it is to simply use AssignGmm_F64 instead.

Specified by:

assign in interface AssignCluster<double[]>

Parameters:

point - Point which is to be assigned

fit - Storage for relative fit quality of each cluster. Length must be at least the number of clusters.

getNumberOfClusters

public int getNumberOfClusters()

Description copied from interface: AssignCluster

Total number of clusters.

Specified by:

getNumberOfClusters in interface AssignCluster<double[]>

Returns:

The total number of clusters.

copy

public AssignCluster<double[]> copy()

Description copied from interface: AssignCluster

Creates an exact copy of this class.

Specified by:

copy in interface AssignCluster<double[]>

Returns:

Copy of class

getClusters

public java.util.List<double[]> getClusters()

setClusters

public void setClusters(java.util.List<double[]> clusters)