This class implements the Unsupervised Backpropagation algorithm, as described in Gashler, Michael S. and Smith, Michael R. and Morris, Richard and Martinez, Tony. Missing Value Imputation With Unsupervised Backpropagation. Computational Intelligence, Wiley Online Library. 2014. This algorithm is very similar to an earlier algorithm called NonlinearPCA, except with the addition of a three-pass training approach that yields better accuracy. If you call noThreePass() before you call train(), then this class implements NonlinearPCA, as published in Scholz, M. Kaplan, F. Guy, C. L. Kopka, J. Selbig, J., Non-linear PCA: a missing data approach, In Bioinformatics, Vol. 21, Number 20, pp. 3887-3895, Oxford University Press, 2005. More...

#include <GRecommender.h>

Inheritance diagram for GClasses::GNonlinearPCA:

Public Member Functions
	GNonlinearPCA (size_t intrinsicDims)
	General-purpose constructor. More...

	GNonlinearPCA (GDomNode *pNode, GLearnerLoader &ll)
	Deserialization constructor. More...

virtual	~GNonlinearPCA ()
	Destructor. More...

void	clampItemElement (size_t item, size_t attr, double val)
	Specify that a certain attribute of a certain item profile has a fixed value. (Values for attr are from 0 to m_pModel->outputLayer().inputs()-1. No mechanism is provided to clamp the item bias.) More...

void	clampItems (const GMatrix &data, size_t offset=0)
	Assumes that column 0 of data is an item ID, and all other columns specify profile values to clamp beginning at the specifed profile offset. More...

void	clampUserElement (size_t user, size_t attr, double val)
	Specify that a certain attribute of a certain user profile has a fixed value. (Values for attr are from 0 to m_intrinsicDims-2. No mechanism is provided to clamp the input bias.) More...

void	clampUsers (const GMatrix &data, size_t offset=0)
	Assumes that column 0 of data is a user ID, and all other columns specify profile values to clamp beginning at the specifed profile offset. More...

virtual void	impute (double *pVec, size_t dims)
	See the comment for GCollaborativeFilter::impute. More...

GNeuralNet *	model ()
	Returns a pointer to the neural net that is used to model the recommendation space. You may want to use this method to add layers to the network. (At least one layer is necessary). You may also use it to set the learning rate, or change activation functions before the model is trained. More...

void	noInputBias ()
	Specify to use no bias value with the inputs. More...

void	noThreePass ()
	Specify not to use three-pass training. (It will just use one pass instead.) More...

virtual double	predict (size_t user, size_t item)
	See the comment for GCollaborativeFilter::predict. More...

virtual GDomNode *	serialize (GDom *pDoc) const
	See the comment for GCollaborativeFilter::serialize. More...

void	setDecayRate (double d)
	Set the rate to decay the learning rate. More...

void	setMinIters (size_t i)
	Sset the min number of iterations to train. More...

void	setRegularizer (double d)
	Set the regularization value. More...

virtual void	train (GMatrix &data)
	See the comment for GCollaborativeFilter::train. More...

GMatrix *	users ()
	Returns a pointer to the matrix of user preference vectors. More...

Public Member Functions inherited from GClasses::GCollaborativeFilter
	GCollaborativeFilter ()

	GCollaborativeFilter (GDomNode *pNode, GLearnerLoader &ll)

virtual	~GCollaborativeFilter ()

void	basicTest (double minMSE)
	Performs a basic unit test on this collaborative filter. More...

double	crossValidate (GMatrix &data, size_t folds, double *pOutMAE=NULL)
	This randomly assigns each rating to one of the folds. Then, for each fold, it calls train with a dataset that contains everything except for the ratings in that fold. It predicts values for the items in the fold, and returns the mean-squared difference between the predictions and the actual ratings. If pOutMAE is non-NULL, it will be set to the mean-absolute error. More...

GMatrix *	precisionRecall (GMatrix &data, bool ideal=false)
	This divides the data into two equal-size parts. It trains on one part, and then measures the precision/recall using the other part. It returns a three-column data set with recall scores in column 0 and corresponding precision scores in column 1. The false-positive rate is in column 2. (So, if you want a precision-recall plot, just drop column 2. If you want an ROC curve, drop column 1 and swap the remaining two columns.) This method assumes the ratings range from 0 to 1, so be sure to scale the ratings to fit that range before calling this method. If ideal is true, then it will ignore your model and report the ideal results as if your model always predicted the correct rating. (This is useful because it shows the best possible results.) More...

GRand &	rand ()
	Returns a reference to the pseudo-random number generator associated with this object. More...

double	trainAndTest (GMatrix &train, GMatrix &test, double *pOutMAE=NULL)
	This trains on the training set, and then tests on the test set. Returns the mean-squared difference between actual and target predictions. More...

void	trainDenseMatrix (const GMatrix &data, const GMatrix *pLabels=NULL)
	Train from an m-by-n dense matrix, where m is the number of users and n is the number of items. All attributes must be continuous. Missing values are indicated with UNKNOWN_REAL_VALUE. If pLabels is non-NULL, then the labels will be appended as additional items. More...

Static Public Member Functions
static void	test ()
	Performs unit tests. Throws if a failure occurs. Returns if successful. More...

Static Public Member Functions inherited from GClasses::GCollaborativeFilter
static double	areaUnderCurve (GMatrix &data)
	Pass in the data returned by the precisionRecall function (unmodified), and this will compute the area under the ROC curve. More...

Protected Member Functions
void	clampItemsInternal (size_t i)

void	clampUsersInternal (size_t i)

double	validate (GNeuralNet *pNN, GMatrix &data)
	Returns the sum-squared error for the specified set of ratings. More...

Protected Member Functions inherited from GClasses::GCollaborativeFilter
GDomNode *	baseDomNode (GDom pDoc, const char szClassName) const
	Child classes should use this in their implementation of serialize. More...

Protected Attributes
double	m_decayRate

size_t	m_intrinsicDims

size_t	m_items

size_t	m_minIters

GMatrix *	m_pItemMask

double *	m_pMaxs

double *	m_pMins

GNeuralNet *	m_pModel

GMatrix *	m_pUserMask

GMatrix *	m_pUsers

double	m_regularizer

bool	m_useInputBias

bool	m_useThreePass

Protected Attributes inherited from GClasses::GCollaborativeFilter
GRand	m_rand

Detailed Description

This class implements the Unsupervised Backpropagation algorithm, as described in Gashler, Michael S. and Smith, Michael R. and Morris, Richard and Martinez, Tony. Missing Value Imputation With Unsupervised Backpropagation. Computational Intelligence, Wiley Online Library. 2014. This algorithm is very similar to an earlier algorithm called NonlinearPCA, except with the addition of a three-pass training approach that yields better accuracy. If you call noThreePass() before you call train(), then this class implements NonlinearPCA, as published in Scholz, M. Kaplan, F. Guy, C. L. Kopka, J. Selbig, J., Non-linear PCA: a missing data approach, In Bioinformatics, Vol. 21, Number 20, pp. 3887-3895, Oxford University Press, 2005.

Constructor & Destructor Documentation

GClasses::GNonlinearPCA::GNonlinearPCA ( size_t intrinsicDims )

General-purpose constructor.

GClasses::GNonlinearPCA::GNonlinearPCA	(	GDomNode *	pNode,
		GLearnerLoader &	ll
	)

Deserialization constructor.

virtual GClasses::GNonlinearPCA::~GNonlinearPCA ( )

virtual

Destructor.

Member Function Documentation

void GClasses::GNonlinearPCA::clampItemElement	(	size_t	item,
		size_t	attr,
		double	val
	)

Specify that a certain attribute of a certain item profile has a fixed value. (Values for attr are from 0 to m_pModel->outputLayer().inputs()-1. No mechanism is provided to clamp the item bias.)

void GClasses::GNonlinearPCA::clampItems	(	const GMatrix &	data,
		size_t	offset = `0`
	)

Assumes that column 0 of data is an item ID, and all other columns specify profile values to clamp beginning at the specifed profile offset.

void GClasses::GNonlinearPCA::clampItemsInternal ( size_t i )

protected

void GClasses::GNonlinearPCA::clampUserElement	(	size_t	user,
		size_t	attr,
		double	val
	)

Specify that a certain attribute of a certain user profile has a fixed value. (Values for attr are from 0 to m_intrinsicDims-2. No mechanism is provided to clamp the input bias.)

void GClasses::GNonlinearPCA::clampUsers	(	const GMatrix &	data,
		size_t	offset = `0`
	)

Assumes that column 0 of data is a user ID, and all other columns specify profile values to clamp beginning at the specifed profile offset.

void GClasses::GNonlinearPCA::clampUsersInternal ( size_t i )

protected

virtual void GClasses::GNonlinearPCA::impute	(	double *	pVec,
		size_t	dims
	)

virtual

See the comment for GCollaborativeFilter::impute.

Implements GClasses::GCollaborativeFilter.

GNeuralNet* GClasses::GNonlinearPCA::model ( )

inline

Returns a pointer to the neural net that is used to model the recommendation space. You may want to use this method to add layers to the network. (At least one layer is necessary). You may also use it to set the learning rate, or change activation functions before the model is trained.

void GClasses::GNonlinearPCA::noInputBias ( )

inline

Specify to use no bias value with the inputs.

void GClasses::GNonlinearPCA::noThreePass ( )

inline

Specify not to use three-pass training. (It will just use one pass instead.)

virtual double GClasses::GNonlinearPCA::predict	(	size_t	user,
		size_t	item
	)

virtual

See the comment for GCollaborativeFilter::predict.

Implements GClasses::GCollaborativeFilter.

Reimplemented in GClasses::GHybridNonlinearPCA.

virtual GDomNode* GClasses::GNonlinearPCA::serialize ( GDom * pDoc ) const

virtual

See the comment for GCollaborativeFilter::serialize.

Implements GClasses::GCollaborativeFilter.

void GClasses::GNonlinearPCA::setDecayRate ( double d )

inline

Set the rate to decay the learning rate.

void GClasses::GNonlinearPCA::setMinIters ( size_t i )

inline

Sset the min number of iterations to train.

void GClasses::GNonlinearPCA::setRegularizer ( double d )

inline

Set the regularization value.

static void GClasses::GNonlinearPCA::test ( )

static

Performs unit tests. Throws if a failure occurs. Returns if successful.

virtual void GClasses::GNonlinearPCA::train ( GMatrix & data )

virtual

See the comment for GCollaborativeFilter::train.

Implements GClasses::GCollaborativeFilter.

Reimplemented in GClasses::GHybridNonlinearPCA.

GMatrix* GClasses::GNonlinearPCA::users ( )

inline

Returns a pointer to the matrix of user preference vectors.

double GClasses::GNonlinearPCA::validate	(	GNeuralNet *	pNN,
		GMatrix &	data
	)

protected

Returns the sum-squared error for the specified set of ratings.

Member Data Documentation

double GClasses::GNonlinearPCA::m_decayRate

protected

size_t GClasses::GNonlinearPCA::m_intrinsicDims

protected

size_t GClasses::GNonlinearPCA::m_items

protected

size_t GClasses::GNonlinearPCA::m_minIters

protected

GMatrix* GClasses::GNonlinearPCA::m_pItemMask

protected

double* GClasses::GNonlinearPCA::m_pMaxs

protected

double* GClasses::GNonlinearPCA::m_pMins

protected

GNeuralNet* GClasses::GNonlinearPCA::m_pModel

protected

GMatrix* GClasses::GNonlinearPCA::m_pUserMask

protected

GMatrix* GClasses::GNonlinearPCA::m_pUsers

protected

double GClasses::GNonlinearPCA::m_regularizer

protected

bool GClasses::GNonlinearPCA::m_useInputBias

protected

bool GClasses::GNonlinearPCA::m_useThreePass

protected

Public Member Functions

Static Public Member Functions

Protected Member Functions

Protected Attributes

Detailed Description

Constructor & Destructor Documentation

Member Function Documentation

Member Data Documentation