Package elki.clustering.kmedoids

Class CLARA<V>

java.lang.Object

elki.clustering.kmedoids.PAM<V>

elki.clustering.kmedoids.CLARA<V>

Type Parameters:

V - Data type

All Implemented Interfaces:

Algorithm, ClusteringAlgorithm<Clustering<MedoidModel>>, KMedoidsClustering<V>

@Reference(authors="L. Kaufman, P. J. Rousseeuw",title="Clustering Large Data Sets",booktitle="Pattern Recognition in Practice",url="https://doi.org/10.1016/B978-0-444-87877-9.50039-X",bibkey="doi:10.1016/B978-0-444-87877-9.50039-X") @Reference(authors="L. Kaufman, P. J. Rousseeuw",title="Clustering Large Applications (Program CLARA)",booktitle="Finding Groups in Data: An Introduction to Cluster Analysis",url="https://doi.org/10.1002/9780470316801.ch3",bibkey="doi:10.1002/9780470316801.ch3")
public class CLARA<V>
extends PAM<V>

Clustering Large Applications (CLARA) is a clustering method for large data sets based on PAM, partitioning around medoids (PAM) based on sampling.

TODO: use a triangular distance matrix, rather than a hash-map based cache, for a bit better performance and less memory.

Reference:

L. Kaufman, P. J. Rousseeuw
Clustering Large Data Sets
Pattern Recognition in Practice

L. Kaufman, P. J. Rousseeuw
Clustering Large Applications (Program CLARA)
Finding Groups in Data: An Introduction to Cluster Analysis

Since:

0.7.0

Author:

Erich Schubert

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
protected static class	CLARA.CachedDistanceQuery<V>	Cached distance query.
static class	CLARA.Par<V>	Parameterization class.

Nested classes/interfaces inherited from class elki.clustering.kmedoids.PAM

PAM.Instance

Nested classes/interfaces inherited from interface elki.Algorithm

Algorithm.Utils

Field Summary

Fields

Modifier and Type	Field	Description
(package private) boolean	keepmed	Keep the previous medoids in the sample (see page 145).
private static Logging	LOG	Class logger.
(package private) int	numsamples	Number of samples to draw (i.e. iterations).
(package private) RandomFactory	random	Random factory for initialization.
(package private) double	sampling	Sampling rate.

Fields inherited from class elki.clustering.kmedoids.PAM

distance, initializer, k, maxiter

Constructor Summary

Constructors

Constructor	Description
CLARA(Distance<? super V> distance, int k, int maxiter, KMedoidsInitialization<V> initializer, int numsamples, double sampling, boolean keepmed, RandomFactory random)	Constructor.

Method Summary

Modifier and Type	Method	Description
protected static double	assignRemainingToNearestCluster(ArrayDBIDs means, DBIDs ids, DBIDs rids, WritableIntegerDataStore assignment, DistanceQuery<?> distQ)	Returns a list of clusters.
(package private) static DBIDs	randomSample(DBIDs ids, int samplesize, java.util.Random rnd, DBIDs previous)	Draw a random sample of the desired size.
Clustering<MedoidModel>	run(Relation<V> relation)	Run k-medoids clustering.
Clustering<MedoidModel>	run(Relation<V> relation, int k, DistanceQuery<? super V> distQ)	Run k-medoids clustering with a given distance query. Not a very elegant API, but needed for some types of nested k-medoids.

Methods inherited from class elki.clustering.kmedoids.PAM

getInputTypeRestriction, getLogger, initialMedoids, wrapResult

Methods inherited from class java.lang.Object

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

Methods inherited from interface elki.clustering.ClusteringAlgorithm

autorun

Field Detail

LOG

private static final Logging LOG

Class logger.

sampling

double sampling

Sampling rate. If less than 1, it is considered to be a relative value.

numsamples

int numsamples

Number of samples to draw (i.e. iterations).

keepmed

boolean keepmed

Keep the previous medoids in the sample (see page 145).

random

RandomFactory random

Random factory for initialization.

Constructor Detail

CLARA

public CLARA(Distance<? super V> distance,
             int k,
             int maxiter,
             KMedoidsInitialization<V> initializer,
             int numsamples,
             double sampling,
             boolean keepmed,
             RandomFactory random)

Constructor.

Parameters:

distance - Distance function to use

k - Number of clusters to produce

maxiter - Maximum number of iterations

initializer - Initialization function

numsamples - Number of samples (sampling iterations)

sampling - Sampling rate (absolute or relative)

keepmed - Keep the previous medoids in the next sample

random - Random generator

Method Detail

run

public Clustering<MedoidModel> run(Relation<V> relation)

Description copied from interface: KMedoidsClustering

Run k-medoids clustering.

Specified by:

run in interface KMedoidsClustering<V>

Overrides:

run in class PAM<V>

Parameters:

relation - relation to use

Returns:

result

run

public Clustering<MedoidModel> run(Relation<V> relation,
                                   int k,
                                   DistanceQuery<? super V> distQ)

Description copied from interface: KMedoidsClustering

Run k-medoids clustering with a given distance query.
Not a very elegant API, but needed for some types of nested k-medoids.

Specified by:

run in interface KMedoidsClustering<V>

Overrides:

run in class PAM<V>

Parameters:

relation - relation to use

k - Number of clusters

distQ - Distance query to use

Returns:

result

randomSample

static DBIDs randomSample(DBIDs ids,
                          int samplesize,
                          java.util.Random rnd,
                          DBIDs previous)

Draw a random sample of the desired size.

Parameters:

ids - IDs to sample from

samplesize - Sample size

rnd - Random generator

previous - Previous medoids to always include in the sample.

Returns:

Sample

assignRemainingToNearestCluster

protected static double assignRemainingToNearestCluster(ArrayDBIDs means,
                                                        DBIDs ids,
                                                        DBIDs rids,
                                                        WritableIntegerDataStore assignment,
                                                        DistanceQuery<?> distQ)

Returns a list of clusters. The k^th cluster contains the ids of those FeatureVectors, that are nearest to the k^th mean.

Parameters:

means - Object centroids

ids - Object ids

rids - Sample that was already assigned

assignment - cluster assignment

distQ - distance query

Returns:

Sum of distances.