edu.ucla.sspace.matrix.factorization

Class NonNegativeMatrixFactorizationMultiplicative

java.lang.Object

edu.ucla.sspace.matrix.factorization.NonNegativeMatrixFactorizationMultiplicative

All Implemented Interfaces:

MatrixFactorization

public class NonNegativeMatrixFactorizationMultiplicative
extends Object
implements MatrixFactorization

The most basic Non-Negative Matrix Factorization implementation. This implementation uses the original multiplicitive update rules for the two decomposed matrices. If we define our data Matrix as A, then A can be approximated by
WH
W and H are initialized with random uniform values from 0 to 1. They are both updated iteratively using the following rules:
W <- W .* (AHt) ./ (W(HHt)) H <- H .* (WtA) ./ ((WtW)H)
Where t stands for the transpose.

These updates are made by fixing one of the matrices and updating the other for a small number of iterations and then updating the other while keeping the first fixed. This alternating update is done for a large number of iterations.

Author:

Keith Stevens

Field Summary

Fields

Modifier and Type	Field and Description
static String	DEFAULT_INNER_ITERATIONS By default, use this many inner iterations.
static String	DEFAULT_OUTER_ITERATIONS By default, use this many outer iterations.
static String	INNER_ITERATIONS The System Property used to define the number inner multiplicative iterations to use when decomposing a matrix.
static String	OUTER_ITERATIONS The System Property used to define the number outer multiplicative iterations to use when decomposing a matrix.
static String	PROPERTY_PREFIX The base property prefix.

Constructor Summary

Constructors

Constructor and Description
NonNegativeMatrixFactorizationMultiplicative() Creates a new NonNegativeMatrixFactorizationMultiplicative using the system defined Properties
NonNegativeMatrixFactorizationMultiplicative(int innerLoop, int outerLoop) Creates a new NonNegativeMatrixFactorizationMultiplicative.
NonNegativeMatrixFactorizationMultiplicative(Properties props) Creates a new NonNegativeMatrixFactorizationMultiplicative using the provided Properties

Method Summary

Methods

Modifier and Type	Method and Description
Matrix	classFeatures() Returns the latent class by feature Matrix.
Matrix	dataClasses() Returns the data point by latent class Matrix.
void	factorize(MatrixFile mFile, int numDimensions) Factorizes the MatrixFile m.
void	factorize(SparseMatrix matrix, int numDimensions) Factorizes the SparseMatrix m.
MatrixBuilder	getBuilder() Returns a MatrixBuilder that is optimized for MatrixFiles that will be processed by this MatrixFactorization algorithm.
static void	initialize(Matrix m) Initializes every value in m to be a random value between 0 and 1, inclusive.

Methods inherited from class java.lang.Object

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

Field Detail

PROPERTY_PREFIX

public static final String PROPERTY_PREFIX

The base property prefix.

See Also:

Constant Field Values

OUTER_ITERATIONS

public static final String OUTER_ITERATIONS

The System Property used to define the number outer multiplicative iterations to use when decomposing a matrix.

See Also:

Constant Field Values

INNER_ITERATIONS

public static final String INNER_ITERATIONS

The System Property used to define the number inner multiplicative iterations to use when decomposing a matrix.

See Also:

Constant Field Values

DEFAULT_OUTER_ITERATIONS

public static final String DEFAULT_OUTER_ITERATIONS

By default, use this many outer iterations.

See Also:

Constant Field Values

DEFAULT_INNER_ITERATIONS

public static final String DEFAULT_INNER_ITERATIONS

By default, use this many inner iterations.

See Also:

Constant Field Values

Constructor Detail

NonNegativeMatrixFactorizationMultiplicative

public NonNegativeMatrixFactorizationMultiplicative()

Creates a new NonNegativeMatrixFactorizationMultiplicative using the system defined Properties

NonNegativeMatrixFactorizationMultiplicative

public NonNegativeMatrixFactorizationMultiplicative(Properties props)

Creates a new NonNegativeMatrixFactorizationMultiplicative using the provided Properties

NonNegativeMatrixFactorizationMultiplicative

public NonNegativeMatrixFactorizationMultiplicative(int innerLoop,
                                            int outerLoop)

Creates a new NonNegativeMatrixFactorizationMultiplicative.

Parameters:

innerLoop - The number of iterations for updating a single matrix while keeping the matrix fixed

outerloop - The number of innerLoop iterations to perform

Method Detail

factorize

public void factorize(MatrixFile mFile,
             int numDimensions)

Factorizes the MatrixFile m. If m is of size MxN, this will generate two matrices: one of size MxnumDimensions and one of size numDimensionsxN. This method must be called before calling dataClasses and classFeatures.

Specified by:

factorize in interface MatrixFactorization

factorize

public void factorize(SparseMatrix matrix,
             int numDimensions)

Factorizes the SparseMatrix m. If m is of size MxN, this will generate two matrices: one of size MxnumDimensions and one of size numDimensionsxN. This method must be called before calling dataClasses and classFeatures.

Specified by:

factorize in interface MatrixFactorization

dataClasses

public Matrix dataClasses()

Returns the data point by latent class Matrix. This matrix represents the degree by which each data point can be explained by the discovered latent classes. Actual interpretations of the latent classes depends on the actual algorithm used. This Matrix can be used as a reduced representation of the data points themselves.

Specified by:

dataClasses in interface MatrixFactorization

classFeatures

public Matrix classFeatures()

Returns the latent class by feature Matrix. This matrix represents the degree by which each feature affects each latent class. Actual interpretations of this interaction depends on the actual algorithm used. This Matrix can be used as a reduced representation of the features.

Specified by:

classFeatures in interface MatrixFactorization

getBuilder

public MatrixBuilder getBuilder()

Returns a MatrixBuilder that is optimized for MatrixFiles that will be processed by this MatrixFactorization algorithm.

Specified by:

getBuilder in interface MatrixFactorization

initialize

public static void initialize(Matrix m)

Initializes every value in m to be a random value between 0 and 1, inclusive.