io.kgraph.library.cf

Class Sgd

java.lang.Object

io.kgraph.library.cf.Sgd

All Implemented Interfaces:

ComputeFunction<CfLongId,org.jblas.FloatMatrix,Float,FloatMatrixMessage>

public class Sgd
extends Object
implements ComputeFunction<CfLongId,org.jblas.FloatMatrix,Float,FloatMatrixMessage>

Stochastic Gradient Descent (SGD) implementation.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
class	Sgd.InitItemsComputation This computation class is used to initialize the factors of the item nodes in the second superstep.
class	Sgd.InitUsersComputation This computation class is used to initialize the factors of the user nodes in the very first superstep, and send the first updates to the item nodes.

Nested classes/interfaces inherited from interface io.kgraph.pregel.ComputeFunction

ComputeFunction.Aggregators, ComputeFunction.Callback<K,VV,EV,Message>, ComputeFunction.InitCallback, ComputeFunction.MasterCallback, ComputeFunction.ReadAggregators, ComputeFunction.ReadWriteAggregators

Field Summary

Fields

Modifier and Type	Field and Description
static String	GAMMA Keyword for parameter setting the learning rate GAMMA.
static float	GAMMA_DEFAULT Default value for GAMMA.
static String	ITERATIONS Keyword for parameter setting the number of iterations.
static int	ITERATIONS_DEFAULT Default value for ITERATIONS.
static String	LAMBDA Keyword for parameter setting the regularization parameter LAMBDA.
static float	LAMBDA_DEFAULT Default value for LABDA.
static String	MAX_RATING Max rating.
static float	MAX_RATING_DEFAULT Default maximum rating
protected float	maxRating
static String	MIN_RATING Min rating.
static float	MIN_RATING_DEFAULT Default minimum rating
protected float	minRating
static String	RANDOM_SEED Random seed.
static Long	RANDOM_SEED_DEFAULT Default random seed
static String	RMSE_AGGREGATOR Aggregator used to compute the RMSE
static String	RMSE_TARGET Keyword for RMSE aggregator tolerance.
static float	RMSE_TARGET_DEFAULT Default value for parameter enabling the RMSE aggregator.
static String	TOLERANCE Keyword for parameter setting the convergence tolerance
static float	TOLERANCE_DEFAULT Default value for TOLERANCE.
static String	VECTOR_SIZE Keyword for parameter setting the Latent Vector Size.
static int	VECTOR_SIZE_DEFAULT Default value for GAMMA.

Constructor Summary

Constructors

Constructor and Description
Sgd()

Method Summary

Modifier and Type	Method and Description
void	compute(int superstep, VertexWithValue<CfLongId,org.jblas.FloatMatrix> vertex, Iterable<FloatMatrixMessage> messages, Iterable<EdgeWithValue<CfLongId,Float>> edges, ComputeFunction.Callback<CfLongId,org.jblas.FloatMatrix,Float,FloatMatrixMessage> cb) The function for computing a new vertex value or sending messages to the next superstep.
protected long	getTotalNumEdges(ComputeFunction.ReadAggregators aggregators)
void	init(Map<String,?> configs, ComputeFunction.InitCallback cb) Initialize the ComputeFunction, this is the place to register aggregators.
void	masterCompute(int superstep, ComputeFunction.MasterCallback cb) A function for performing sequential computations between supersteps.
void	preSuperstep(int superstep, ComputeFunction.Aggregators aggregators) Prepare for computation.
void	superstepCompute(int superstep, VertexWithValue<CfLongId,org.jblas.FloatMatrix> vertex, Iterable<FloatMatrixMessage> messages, Iterable<EdgeWithValue<CfLongId,Float>> edges, ComputeFunction.Callback<CfLongId,org.jblas.FloatMatrix,Float,FloatMatrixMessage> cb) Main SGD compute method.
protected void	updateValue(org.jblas.FloatMatrix value, org.jblas.FloatMatrix update, float rating, float minRating, float maxRating, float lambda, float gamma) Applies the SGD update logic in the provided vector.

Methods inherited from class java.lang.Object

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

Methods inherited from interface io.kgraph.pregel.ComputeFunction

postSuperstep

Field Detail

RMSE_TARGET

public static final String RMSE_TARGET

Keyword for RMSE aggregator tolerance.

See Also:

Constant Field Values

RMSE_TARGET_DEFAULT

public static final float RMSE_TARGET_DEFAULT

Default value for parameter enabling the RMSE aggregator.

See Also:

Constant Field Values

TOLERANCE

public static final String TOLERANCE

Keyword for parameter setting the convergence tolerance

See Also:

Constant Field Values

TOLERANCE_DEFAULT

public static final float TOLERANCE_DEFAULT

Default value for TOLERANCE.

See Also:

Constant Field Values

ITERATIONS

public static final String ITERATIONS

Keyword for parameter setting the number of iterations.

See Also:

Constant Field Values

ITERATIONS_DEFAULT

public static final int ITERATIONS_DEFAULT

Default value for ITERATIONS.

See Also:

Constant Field Values

LAMBDA

public static final String LAMBDA

Keyword for parameter setting the regularization parameter LAMBDA.

See Also:

Constant Field Values

LAMBDA_DEFAULT

public static final float LAMBDA_DEFAULT

Default value for LABDA.

See Also:

Constant Field Values

GAMMA

public static final String GAMMA

Keyword for parameter setting the learning rate GAMMA.

See Also:

Constant Field Values

GAMMA_DEFAULT

public static final float GAMMA_DEFAULT

Default value for GAMMA.

See Also:

Constant Field Values

VECTOR_SIZE

public static final String VECTOR_SIZE

Keyword for parameter setting the Latent Vector Size.

See Also:

Constant Field Values

VECTOR_SIZE_DEFAULT

public static final int VECTOR_SIZE_DEFAULT

Default value for GAMMA.

See Also:

Constant Field Values

MAX_RATING

public static final String MAX_RATING

Max rating.

See Also:

Constant Field Values

MAX_RATING_DEFAULT

public static final float MAX_RATING_DEFAULT

Default maximum rating

See Also:

Constant Field Values

MIN_RATING

public static final String MIN_RATING

Min rating.

See Also:

Constant Field Values

MIN_RATING_DEFAULT

public static final float MIN_RATING_DEFAULT

Default minimum rating

See Also:

Constant Field Values

RANDOM_SEED

public static final String RANDOM_SEED

Random seed.

See Also:

Constant Field Values

RANDOM_SEED_DEFAULT

public static final Long RANDOM_SEED_DEFAULT

Default random seed

RMSE_AGGREGATOR

public static final String RMSE_AGGREGATOR

Aggregator used to compute the RMSE

See Also:

Constant Field Values

minRating

protected float minRating

maxRating

protected float maxRating

Constructor Detail

Sgd

public Sgd()

Method Detail

preSuperstep

public void preSuperstep(int superstep,
                         ComputeFunction.Aggregators aggregators)

Description copied from interface: ComputeFunction

Prepare for computation. This method is executed exactly once prior to compute() being called for any of the vertices in the partition.

Specified by:

preSuperstep in interface ComputeFunction<CfLongId,org.jblas.FloatMatrix,Float,FloatMatrixMessage>

Parameters:

superstep - the superstep

aggregators - the aggregators

superstepCompute

public void superstepCompute(int superstep,
                             VertexWithValue<CfLongId,org.jblas.FloatMatrix> vertex,
                             Iterable<FloatMatrixMessage> messages,
                             Iterable<EdgeWithValue<CfLongId,Float>> edges,
                             ComputeFunction.Callback<CfLongId,org.jblas.FloatMatrix,Float,FloatMatrixMessage> cb)

Main SGD compute method.

Parameters:

superstep - the count of the current superstep

vertex - the current vertex with its value

messages - a Map of the source vertex and the message sent from the previous superstep

edges - the adjacent edges with their values

cb - a callback for setting a new vertex value or sending messages to the next superstep

updateValue

protected final void updateValue(org.jblas.FloatMatrix value,
                                 org.jblas.FloatMatrix update,
                                 float rating,
                                 float minRating,
                                 float maxRating,
                                 float lambda,
                                 float gamma)

Applies the SGD update logic in the provided vector. It does the update in-place.

The update is performed according to the following formula:

v = v - gamma*(lambda*v + error*u)

Parameters:

value - the vector to update

update - the vector used to update

rating - the rating

minRating - the min rating

maxRating - the max rating

lambda - the lambda parameter

gamma - the gamma parameter

init

public final void init(Map<String,?> configs,
                       ComputeFunction.InitCallback cb)

Description copied from interface: ComputeFunction

Initialize the ComputeFunction, this is the place to register aggregators.

Specified by:

init in interface ComputeFunction<CfLongId,org.jblas.FloatMatrix,Float,FloatMatrixMessage>

Parameters:

configs - configuration parameters

cb - a callback for registering aggregators

masterCompute

public final void masterCompute(int superstep,
                                ComputeFunction.MasterCallback cb)

Description copied from interface: ComputeFunction

A function for performing sequential computations between supersteps.

Specified by:

masterCompute in interface ComputeFunction<CfLongId,org.jblas.FloatMatrix,Float,FloatMatrixMessage>

Parameters:

superstep - the superstep

cb - a callback for writing to aggregators or halting the computation

compute

public void compute(int superstep,
                    VertexWithValue<CfLongId,org.jblas.FloatMatrix> vertex,
                    Iterable<FloatMatrixMessage> messages,
                    Iterable<EdgeWithValue<CfLongId,Float>> edges,
                    ComputeFunction.Callback<CfLongId,org.jblas.FloatMatrix,Float,FloatMatrixMessage> cb)

Description copied from interface: ComputeFunction

The function for computing a new vertex value or sending messages to the next superstep.

Specified by:

compute in interface ComputeFunction<CfLongId,org.jblas.FloatMatrix,Float,FloatMatrixMessage>

Parameters:

superstep - the count of the current superstep

vertex - the current vertex with its value

messages - a Map of the source vertex and the message sent from the previous superstep

edges - the adjacent edges with their values

cb - a callback for setting a new vertex value or sending messages to the next superstep

getTotalNumEdges

protected long getTotalNumEdges(ComputeFunction.ReadAggregators aggregators)