mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy > Class Template Reference

This class implements K-Means clustering. More...

List of all members.

Public Member Functions

 KMeans (const size_t maxIterations=1000, const double overclusteringFactor=1.0, const MetricType metric=MetricType(), const InitialPartitionPolicy partitioner=InitialPartitionPolicy(), const EmptyClusterPolicy emptyClusterAction=EmptyClusterPolicy())
 Create a K-Means object and (optionally) set the parameters which K-Means will be run with.
template<typename MatType >
void Cluster (const MatType &data, const size_t clusters, arma::Col< size_t > &assignments, MatType &centroids, const bool initialAssignmentGuess=false, const bool initialCentroidGuess=false) const
 Perform k-means clustering on the data, returning a list of cluster assignments and also the centroids of each cluster.
template<typename MatType >
void Cluster (const MatType &data, const size_t clusters, arma::Col< size_t > &assignments, const bool initialGuess=false) const
 Perform k-means clustering on the data, returning a list of cluster assignments.
EmptyClusterPolicy & EmptyClusterAction ()
 Modify the empty cluster policy.
const EmptyClusterPolicy & EmptyClusterAction () const
 Get the empty cluster policy.
template<typename MatType >
void FastCluster (MatType &data, const size_t clusters, arma::Col< size_t > &assignments) const
 An implementation of k-means using the Pelleg-Moore algorithm; this is known to not work -- do not use it! (Fixing it is TODO, of course; see #251.
size_t & MaxIterations ()
 Set the maximum number of iterations.
size_t MaxIterations () const
 Get the maximum number of iterations.
MetricType & Metric ()
 Modify the distance metric.
const MetricType & Metric () const
 Get the distance metric.
double & OverclusteringFactor ()
 Set the overclustering factor. Must be greater than 1.
double OverclusteringFactor () const
 Return the overclustering factor.
InitialPartitionPolicy & Partitioner ()
 Modify the initial partitioning policy.
const InitialPartitionPolicy & Partitioner () const
 Get the initial partitioning policy.

Private Attributes

EmptyClusterPolicy emptyClusterAction
 Instantiated empty cluster policy.
size_t maxIterations
 Maximum number of iterations before giving up.
MetricType metric
 Instantiated distance metric.
double overclusteringFactor
 Factor controlling how many clusters are actually found.
InitialPartitionPolicy partitioner
 Instantiated initial partitioning policy.

Detailed Description

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
class mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >

This class implements K-Means clustering.

This implementation supports overclustering, which means that more clusters than are requested will be found; then, those clusters will be merged together to produce the desired number of clusters.

Two template parameters can (optionally) be supplied: the policy for how to find the initial partition of the data, and the actions to be taken when an empty cluster is encountered, as well as the distance metric to be used.

A simple example of how to run K-Means clustering is shown below.

 extern arma::mat data; // Dataset we want to run K-Means on.
 arma::Col<size_t> assignments; // Cluster assignments.

 KMeans<> k; // Default options.
 k.Cluster(data, 3, assignments); // 3 clusters.

 // Cluster using the Manhattan distance, 100 iterations maximum, and an
 // overclustering factor of 4.0.
 KMeans<metric::ManhattanDistance> k(100, 4.0);
 k.Cluster(data, 6, assignments); // 6 clusters.
Template Parameters:
MetricType The distance metric to use for this KMeans; see metric::LMetric for an example.
InitialPartitionPolicy Initial partitioning policy; must implement a default constructor and 'void Cluster(const arma::mat&, const size_t, arma::Col<size_t>&)'.
EmptyClusterPolicy Policy for what to do on an empty cluster; must implement a default constructor and 'void EmptyCluster(const arma::mat&, arma::Col<size_t&)'.
See also:
RandomPartition, RefinedStart, AllowEmptyClusters, MaxVarianceNewCluster

Definition at line 75 of file kmeans.hpp.


Constructor & Destructor Documentation

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::KMeans ( const size_t  maxIterations = 1000,
const double  overclusteringFactor = 1.0,
const MetricType  metric = MetricType(),
const InitialPartitionPolicy  partitioner = InitialPartitionPolicy(),
const EmptyClusterPolicy  emptyClusterAction = EmptyClusterPolicy() 
)

Create a K-Means object and (optionally) set the parameters which K-Means will be run with.

This implementation allows a few strategies to improve the performance of K-Means, including "overclustering" and disallowing empty clusters.

The overclustering factor controls how many clusters are actually found; for instance, with an overclustering factor of 4, if K-Means is run to find 3 clusters, it will actually find 12, then merge the nearest clusters until only 3 are left.

Parameters:
maxIterations Maximum number of iterations allowed before giving up (0 is valid, but the algorithm may never terminate).
overclusteringFactor Factor controlling how many extra clusters are found and then merged to get the desired number of clusters.
metric Optional MetricType object; for when the metric has state it needs to store.
partitioner Optional InitialPartitionPolicy object; for when a specially initialized partitioning policy is required.
emptyClusterAction Optional EmptyClusterPolicy object; for when a specially initialized empty cluster policy is required.

Member Function Documentation

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
template<typename MatType >
void mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::Cluster ( const MatType &  data,
const size_t  clusters,
arma::Col< size_t > &  assignments,
MatType &  centroids,
const bool  initialAssignmentGuess = false,
const bool  initialCentroidGuess = false 
) const [inline]

Perform k-means clustering on the data, returning a list of cluster assignments and also the centroids of each cluster.

Optionally, the vector of assignments can be set to an initial guess of the cluster assignments; to do this, set initialAssignmentGuess to true. Another way to set initial cluster guesses is to fill the centroids matrix with the centroid guesses, and then set initialCentroidGuess to true. initialAssignmentGuess supersedes initialCentroidGuess, so if both are set to true, the assignments vector is used.

Note that if the overclustering factor is greater than 1, the centroids matrix will be resized in the method. Regardless of the overclustering factor, the centroid guess matrix (if initialCentroidGuess is set to true) should have the same number of rows as the data matrix, and number of columns equal to 'clusters'.

Template Parameters:
MatType Type of matrix (arma::mat or arma::sp_mat).
Parameters:
data Dataset to cluster.
clusters Number of clusters to compute.
assignments Vector to store cluster assignments in.
centroids Matrix in which centroids are stored.
initialAssignmentGuess If true, then it is assumed that assignments has a list of initial cluster assignments.
initialCentroidGuess If true, then it is assumed that centroids contains the initial centroids of each cluster.
template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
template<typename MatType >
void mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::Cluster ( const MatType &  data,
const size_t  clusters,
arma::Col< size_t > &  assignments,
const bool  initialGuess = false 
) const [inline]

Perform k-means clustering on the data, returning a list of cluster assignments.

Optionally, the vector of assignments can be set to an initial guess of the cluster assignments; to do this, set initialGuess to true.

Template Parameters:
MatType Type of matrix (arma::mat or arma::sp_mat).
Parameters:
data Dataset to cluster.
clusters Number of clusters to compute.
assignments Vector to store cluster assignments in.
initialGuess If true, then it is assumed that assignments has a list of initial cluster assignments.
template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
EmptyClusterPolicy& mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::EmptyClusterAction (  )  [inline]

Modify the empty cluster policy.

Definition at line 194 of file kmeans.hpp.

References mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::emptyClusterAction.

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
const EmptyClusterPolicy& mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::EmptyClusterAction (  )  const [inline]
template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
template<typename MatType >
void mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::FastCluster ( MatType &  data,
const size_t  clusters,
arma::Col< size_t > &  assignments 
) const [inline]

An implementation of k-means using the Pelleg-Moore algorithm; this is known to not work -- do not use it! (Fixing it is TODO, of course; see #251.

)

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
size_t& mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::MaxIterations (  )  [inline]

Set the maximum number of iterations.

Definition at line 178 of file kmeans.hpp.

References mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::maxIterations.

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
size_t mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::MaxIterations (  )  const [inline]

Get the maximum number of iterations.

Definition at line 176 of file kmeans.hpp.

References mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::maxIterations.

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
MetricType& mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::Metric (  )  [inline]
template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
const MetricType& mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::Metric (  )  const [inline]
template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
double& mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::OverclusteringFactor (  )  [inline]

Set the overclustering factor. Must be greater than 1.

Definition at line 173 of file kmeans.hpp.

References mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::overclusteringFactor.

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
double mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::OverclusteringFactor (  )  const [inline]

Return the overclustering factor.

Definition at line 171 of file kmeans.hpp.

References mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::overclusteringFactor.

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
InitialPartitionPolicy& mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::Partitioner (  )  [inline]

Modify the initial partitioning policy.

Definition at line 188 of file kmeans.hpp.

References mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::partitioner.

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
const InitialPartitionPolicy& mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::Partitioner (  )  const [inline]

Get the initial partitioning policy.

Definition at line 186 of file kmeans.hpp.

References mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::partitioner.


Member Data Documentation

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
EmptyClusterPolicy mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::emptyClusterAction [private]

Instantiated empty cluster policy.

Definition at line 206 of file kmeans.hpp.

Referenced by mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::EmptyClusterAction().

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
size_t mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::maxIterations [private]

Maximum number of iterations before giving up.

Definition at line 200 of file kmeans.hpp.

Referenced by mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::MaxIterations().

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
MetricType mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::metric [private]
template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
double mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::overclusteringFactor [private]

Factor controlling how many clusters are actually found.

Definition at line 198 of file kmeans.hpp.

Referenced by mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::OverclusteringFactor().

template<typename MetricType = metric::SquaredEuclideanDistance, typename InitialPartitionPolicy = RandomPartition, typename EmptyClusterPolicy = MaxVarianceNewCluster>
InitialPartitionPolicy mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::partitioner [private]

Instantiated initial partitioning policy.

Definition at line 204 of file kmeans.hpp.

Referenced by mlpack::kmeans::KMeans< MetricType, InitialPartitionPolicy, EmptyClusterPolicy >::Partitioner().


The documentation for this class was generated from the following file:

Generated on 13 Aug 2014 for MLPACK by  doxygen 1.6.1