jsat.datatransform.visualization

Class TSNE

java.lang.Object

jsat.datatransform.visualization.TSNE

All Implemented Interfaces:

Serializable, Cloneable, VisualizationTransform

public class TSNE
extends Object
implements VisualizationTransform

t-distributed Stochastic Neighbor Embedding is an algorithm for creating low dimensional embeddings of datasets, for the purpose of visualization. It attempts to keep points that are near each other in the original space near each other in the low dimensional space as well, with less emphasis on maintaining far-away relationships in the data. This implementation uses the approximated gradients to learn the embedding in O(n log n) time.

If the input dataset has a dimension greater than 50, it is advisable to project the data set down to 50 dimensions using PCA or some similar technique.

See:

• Maaten, L. Van Der, & Hinton, G. (2008). Visualizing Data using t-SNE. Journal of Machine Learning Research, 9, 2579–2605.
• Van der Maaten, L. (2014). Accelerating t-SNE using Tree-Based Algorithms. Journal of Machine Learning Research, 15, 3221–3245. Retrieved from here

Author:

Edward Raff

See Also:

Serialized Form

Constructor Summary

Constructors

Constructor and Description
TSNE()

Method Summary

Modifier and Type	Method and Description
protected static void	computeP(DataSet d, ExecutorService ex, Random rand, int knn, int[][] nearMe, double[][] nearMePij, DistanceMetric dm, double perplexity)
double	getAlpha()
int	getIterations()
double	getPerplexity()
int	getTargetDimension()
void	setAlpha(double alpha) α is the "early exaggeration" constant.
void	setIterations(int T) Sets the desired number of gradient descent iterations to perform.
void	setPerplexity(double perplexity) Sets the target perplexity of the gaussian used over each data point.
boolean	setTargetDimension(int target) Sets the target dimension to embed new dataset to.
<Type extends DataSet> Type	transform(DataSet<Type> d) Transforms the given data set, returning a dataset of the same type.
<Type extends DataSet> Type	transform(DataSet<Type> d, ExecutorService ex) Transforms the given data set, returning a dataset of the same type.

Methods inherited from class java.lang.Object

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

Constructor Detail

TSNE

public TSNE()

Method Detail

setAlpha

public void setAlpha(double alpha)

α is the "early exaggeration" constant. It is a multiple applied to part of the gradient for th first quarter of iterations, and can improve the quality of the solution found. A value in the range of [4, 20] is recommended.

Parameters:

alpha - the exaggeration constant

getAlpha

public double getAlpha()

Returns:

the "early exaggeration" constant

setPerplexity

public void setPerplexity(double perplexity)

Sets the target perplexity of the gaussian used over each data point. The perplexity can be thought of as a quasi desired number of nearest neighbors to be considered, but is adapted based on the distribution of the data. Increasing the perplexity can increase the amount of time it takes to get an embedding. Using a value in the range of [5, 50] is recommended.

Parameters:

perplexity - the quasi number of neighbors to consider for each data point

getPerplexity

public double getPerplexity()

Returns:

the target perplexity to use for each data point

setIterations

public void setIterations(int T)

Sets the desired number of gradient descent iterations to perform.

Parameters:

T - the number of gradient descent iterations

getIterations

public int getIterations()

Returns:

the number of gradient descent iterations to perform

transform

public <Type extends DataSet> Type transform(DataSet<Type> d)

Description copied from interface: VisualizationTransform

Transforms the given data set, returning a dataset of the same type.

Specified by:

transform in interface VisualizationTransform

Type Parameters:

Type - the dataset type

Parameters:

d - the data set to transform

Returns:

the lower dimension dataset for visualization.

transform

public <Type extends DataSet> Type transform(DataSet<Type> d,
                                             ExecutorService ex)

Description copied from interface: VisualizationTransform

Transforms the given data set, returning a dataset of the same type.

Specified by:

transform in interface VisualizationTransform

Type Parameters:

Type - the dataset type

Parameters:

d - the data set to transform

ex - the source of threads for parallel computation

Returns:

the lower dimension dataset for visualization.

computeP

protected static void computeP(DataSet d,
                               ExecutorService ex,
                               Random rand,
                               int knn,
                               int[][] nearMe,
                               double[][] nearMePij,
                               DistanceMetric dm,
                               double perplexity)

Parameters:

d - the dataset to search

ex - the source of threads for parallel computation

rand - source of randomness

knn - the number of neighbors to search for

nearMe - each row is the set of knn indices returned by the NN search

nearMePij - the symmetrized neighbor probability

dm - the distance metric to use for determining closeness

perplexity - the perplexity value for the effective nearest neighbor search and weighting

getTargetDimension

public int getTargetDimension()

Specified by:

getTargetDimension in interface VisualizationTransform

Returns:

the number of dimensions that a dataset will be embedded down to

setTargetDimension

public boolean setTargetDimension(int target)

Description copied from interface: VisualizationTransform

Sets the target dimension to embed new dataset to. Many visualization methods may only support a target of 2 or 3 dimensions, or only one of those options. For that reason a boolean value will be returned indicating if the target size was acceptable. If not, no change to the object will occur.

Specified by:

setTargetDimension in interface VisualizationTransform

Parameters:

target - the new target dimension size when VisualizationTransform.transform(jsat.DataSet) is called.

Returns:

true if this transform supports that dimension and it was set, false if the target dimension is unsupported and the previous value will be used instead.