Package eu.h2020.helios_social.modules.socialgraphmining.GNN

Class GNNMiner

java.lang.Object

eu.h2020.helios_social.modules.socialgraphmining.SocialGraphMiner

eu.h2020.helios_social.modules.socialgraphmining.GNN.GNNMiner

public class GNNMiner
extends SocialGraphMiner

This class provides an implementation of a SocialGraphMiner based on a Graph Neural Network (GNN) architecture. Its parameters can be adjusted using a number of setter methods.

Author:

Emmanouil Krasanakis

Nested Class Summary

Nested classes/interfaces inherited from class eu.h2020.helios_social.modules.socialgraphmining.SocialGraphMiner

SocialGraphMiner.InteractionType, SocialGraphMiner.SocialGraphMinerParameters

Constructor Summary

Constructors

Constructor	Description
GNNMiner(eu.h2020.helios_social.core.contextualegonetwork.ContextualEgoNetwork contextualEgoNetwork)	Instantiates a GNNMiner on a given contextual ego network.

Method Summary

Modifier and Type	Method	Description
SocialGraphMiner.SocialGraphMinerParameters	constructModelParameterObject(eu.h2020.helios_social.core.contextualegonetwork.Interaction interaction)
double	getConfidence(eu.h2020.helios_social.core.contextualegonetwork.Context context)
void	newInteractionParameters(eu.h2020.helios_social.core.contextualegonetwork.Interaction interaction, SocialGraphMiner.SocialGraphMinerParameters params, SocialGraphMiner.InteractionType interactionType)
double	predictNewInteraction(eu.h2020.helios_social.core.contextualegonetwork.Context context, eu.h2020.helios_social.core.contextualegonetwork.Node destinationNode)	Predicts the weight of performing a SEND interaction between the given context's ego and a destination node within a given context.
double	predictNewInteraction(eu.h2020.helios_social.core.contextualegonetwork.Context context, eu.h2020.helios_social.core.contextualegonetwork.Node u, eu.h2020.helios_social.core.contextualegonetwork.Node v)
GNNMiner	setDeniability(double plausibleDeniability, double differentialPrivacy)	Enables plausible deniability and differential privacy handling by permuting the ego and its alter's parameters with a random noise proportional to a given constant and their norm.
GNNMiner	setEdgePointsLearningMultiplier(double incomming, double outgoing)
GNNMiner	setEmbeddingExchangeProtocol(EmbeddingExchangeProtocol embeddingExchangeProtocol)
GNNMiner	setLearningRate(double learningRate)	The learning rate (default is 1) from which GNNMiner training starts.
GNNMiner	setLearningRateDegradation(double learningRateDegradation)	Performs a fixed degradation of the learning rate over training epochs by multiplying the latter with a given factor (default is 0.95) after each epoch.
GNNMiner	setMaxTrainingEpoch(int maxEpoch)	Limits the number of training epochs (default is 1000) over which to train the GNNMiner.
GNNMiner	setMinTrainingRelativeLoss(double convergenceRelativeLoss)	When the GNNMiner is being trained, training stops at epochs where abs(previous epoch loss - this epoch loss) < convergenceRelativeLoss*(this epoch loss) where losses are weighted cross entropy ones.
GNNMiner	setRegularizationAbsorbsion(double regularizationAbsorbsion)	Multiplies regularization tensors with this value before setting them as regularization; value of 1 (default) produces regularization of calculated alter embeddings towards the embeddings calculated on alter devices.
GNNMiner	setRegularizationWeight(double regularizationWeight)	The regularization weight (default 0.1) to apply during training of the GNNMiner.
GNNMiner	setSpectralAlignment(boolean enableSpectralAlignment)	Faster convergence to more robust embeddings of evolving user preferences by trying to align the dimensions of received embeddings towards with their locally understood spectral representation (default is false).
GNNMiner	setTrainingExampleDegradation(double trainingExampleDegradation)	Degrades example weights each time a new one is generated through SocialGraphMiner.newInteraction(eu.h2020.helios_social.core.contextualegonetwork.Interaction, java.lang.String, eu.h2020.helios_social.modules.socialgraphmining.SocialGraphMiner.InteractionType) by calling ContextTrainingExampleData.degrade(double, double) to multiply previous weights with the given degradation factor (default is 0.5).
GNNMiner	setTrainingExamplePropagation(boolean enableTrainingExamplePropagation)	Sets whether to propagate training examples to communicating devices (this behavior is de-activeated by default).
GNNMiner	setTrainingExampleRemovalThreshold(double trainingExampleRemovalThreshold)	Sets the threshold weight at which old training examples are removed (default is 0.1).

Methods inherited from class java.lang.Object

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

Methods inherited from class eu.h2020.helios_social.modules.socialgraphmining.SocialGraphMiner

getContextualEgoNetwork, getModelParameterObject, getModelParameters, newInteraction, recommendInteractions, setSendPermision

Constructor Detail

GNNMiner

public GNNMiner(eu.h2020.helios_social.core.contextualegonetwork.ContextualEgoNetwork contextualEgoNetwork)

Instantiates a GNNMiner on a given contextual ego network.

Parameters:

contextualEgoNetwork - The contextual ego network on which the miner runs and stores information.

Method Detail

setLearningRate

public GNNMiner setLearningRate(double learningRate)

The learning rate (default is 1) from which GNNMiner training starts. Training restarts on each SocialGraphMiner.newInteraction(eu.h2020.helios_social.core.contextualegonetwork.Interaction, java.lang.String, eu.h2020.helios_social.modules.socialgraphmining.SocialGraphMiner.InteractionType) from this value and can be potentially be adapted by setLearningRateDegradation(double) over training epochs.

Parameters:

learningRate - The learning rate to restart from. Should be positive.

Returns:

this GNNMiner instance.

See Also:

GNNNodeData.setLearningRate(double)

setEmbeddingExchangeProtocol

public GNNMiner setEmbeddingExchangeProtocol(EmbeddingExchangeProtocol embeddingExchangeProtocol)

setEdgePointsLearningMultiplier

public GNNMiner setEdgePointsLearningMultiplier(double incomming,
                                                double outgoing)

setLearningRateDegradation

public GNNMiner setLearningRateDegradation(double learningRateDegradation)

Performs a fixed degradation of the learning rate over training epochs by multiplying the latter with a given factor (default is 0.95) after each epoch.

Parameters:

learningRateDegradation - The rate at which learning rate degrade. Should lie in the range (0,1].

Returns:

this GNNMiner instance.

See Also:

setLearningRate(double)

setRegularizationWeight

public GNNMiner setRegularizationWeight(double regularizationWeight)

The regularization weight (default 0.1) to apply during training of the GNNMiner. This weight ensures that training converges around given areas of the embedding space. Value of 0 disables regularization altogether, but does not guarantee convergence (prefer setting setRegularizationAbsorbsion(0) to disable regularization propagation instead).

Parameters:

regularizationWeight - The regularization weight to set.

Returns:

this GNNMiner instance.

See Also:

GNNNodeData.setRegularizationWeight(double), setRegularizationAbsorbsion(double)

setRegularizationAbsorbsion

public GNNMiner setRegularizationAbsorbsion(double regularizationAbsorbsion)

Multiplies regularization tensors with this value before setting them as regularization; value of 1 (default) produces regularization of calculated alter embeddings towards the embeddings calculated on alter devices. Value of 0 produce regularization towards zero, which effectively limits the embedding norm to the approximate order of magnitude 1/weight, where weight is the value set to setRegularizationWeight(double).

Parameters:

regularizationAbsorbsion - A value in the range [0,1].

Returns:

this GNNMiner instance.

setMaxTrainingEpoch

public GNNMiner setMaxTrainingEpoch(int maxEpoch)

Limits the number of training epochs (default is 1000) over which to train the GNNMiner.

Parameters:

maxEpoch - The maximum epoch at which to train.

Returns:

this GNNMiner instance.

setMinTrainingRelativeLoss

public GNNMiner setMinTrainingRelativeLoss(double convergenceRelativeLoss)

When the GNNMiner is being trained, training stops at epochs where abs(previous epoch loss - this epoch loss) < convergenceRelativeLoss*(this epoch loss) where losses are weighted cross entropy ones. Default is 0.001.

Parameters:

convergenceRelativeLoss - The relative loss at which to stop training.

Returns:

this GNNMiner instance.

setTrainingExampleDegradation

public GNNMiner setTrainingExampleDegradation(double trainingExampleDegradation)

Degrades example weights each time a new one is generated through SocialGraphMiner.newInteraction(eu.h2020.helios_social.core.contextualegonetwork.Interaction, java.lang.String, eu.h2020.helios_social.modules.socialgraphmining.SocialGraphMiner.InteractionType) by calling ContextTrainingExampleData.degrade(double, double) to multiply previous weights with the given degradation factor (default is 0.5).

Parameters:

trainingExampleDegradation - The factor with which to multiply each 8previous example weight

Returns:

this GNNMiner instance.

See Also:

setTrainingExampleRemovalThreshold(double)

setTrainingExampleRemovalThreshold

public GNNMiner setTrainingExampleRemovalThreshold(double trainingExampleRemovalThreshold)

Sets the threshold weight at which old training examples are removed (default is 0.1). Basically, if the degradation set by set by setTrainingExampleDegradation(double) remains constant throughout training iterations, training examples are removed if degradation^n < trainingExampleRemovalThreshold where n the number of (positive) examples provided after the examined one with SocialGraphMiner.newInteraction(eu.h2020.helios_social.core.contextualegonetwork.Interaction, java.lang.String, eu.h2020.helios_social.modules.socialgraphmining.SocialGraphMiner.InteractionType).

Parameters:

trainingExampleRemovalThreshold - The weight threshold at which to remove GNN training examples which is passed as the second argument to ContextTrainingExampleData.degrade(double, double) calls.

Returns:

this GNNMiner instance.

setDeniability

public GNNMiner setDeniability(double plausibleDeniability,
                               double differentialPrivacy)

Enables plausible deniability and differential privacy handling by permuting the ego and its alter's parameters with a random noise proportional to a given constant and their norm. Zero values (default) ensure no privacy concerns but more exact computations for other devices. The user's device would perform predictions depending on the privacy settings of their alters.

Parameters:

plausibleDeniability - The permutation of the ego's parameters.

differentialPrivacy - The permutation of the neighbor's parameters.

Returns:

this GNNMiner instance.

setTrainingExamplePropagation

public GNNMiner setTrainingExamplePropagation(boolean enableTrainingExamplePropagation)

Sets whether to propagate training examples to communicating devices (this behavior is de-activeated by default). Propagated examples contribute to the training of the neighbors. This partially violates the device's and all of its neighbors' privacy regardless of what values are set with setDeniability(double, double).

Parameters:

enableTrainingExamplePropagation - Whether to propagate training examples or not.

Returns:

this GNNMiner instance.

setSpectralAlignment

public GNNMiner setSpectralAlignment(boolean enableSpectralAlignment)

Faster convergence to more robust embeddings of evolving user preferences by trying to align the dimensions of received embeddings towards with their locally understood spectral representation (default is false).

Parameters:

enableSpectralAlignment - Whether to enable spectral alignment of learned embeddings between devices.

Returns:

this GNNMiner instance.

newInteractionParameters

public void newInteractionParameters(eu.h2020.helios_social.core.contextualegonetwork.Interaction interaction,
                                     SocialGraphMiner.SocialGraphMinerParameters params,
                                     SocialGraphMiner.InteractionType interactionType)

Specified by:

newInteractionParameters in class SocialGraphMiner

getConfidence

public double getConfidence(eu.h2020.helios_social.core.contextualegonetwork.Context context)

constructModelParameterObject

public SocialGraphMiner.SocialGraphMinerParameters constructModelParameterObject(eu.h2020.helios_social.core.contextualegonetwork.Interaction interaction)

predictNewInteraction

public double predictNewInteraction(eu.h2020.helios_social.core.contextualegonetwork.Context context,
                                    eu.h2020.helios_social.core.contextualegonetwork.Node u,
                                    eu.h2020.helios_social.core.contextualegonetwork.Node v)

predictNewInteraction

public double predictNewInteraction(eu.h2020.helios_social.core.contextualegonetwork.Context context,
                                    eu.h2020.helios_social.core.contextualegonetwork.Node destinationNode)

Description copied from class: SocialGraphMiner

Predicts the weight of performing a SEND interaction between the given context's ego and a destination node within a given context. This method should typically return values in the [0,1] range, where higher values correspond to more likely interactions.

Specified by:

predictNewInteraction in class SocialGraphMiner

Parameters:

context - The context in which to perform the prediction.

destinationNode - The destination node of the interaction.

Returns:

The weight of performing the interaction (higher values are more likely).