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

Class GNNNodeData

java.lang.Object

eu.h2020.helios_social.modules.socialgraphmining.GNN.GNNNodeData

public class GNNNodeData
extends java.lang.Object

This class provides a storage structure that organizes an embedding and a regularization target of contextual ego network nodes. It is indented to be used as a dynamically created instance on nodes (which are cross module components) by calling node.getOrCreateInstance(GNNNodeData.class) to either retrieve of create it.

Author:

Emmanouil Krasanakis

Constructor Summary

Constructors

Constructor	Description
GNNNodeData()

Method Summary

Modifier and Type	Method	Description
void	addEmbeddingToHistory()
void	addRegularizationToHistory()
void	forceSetEmbedding(mklab.JGNN.core.Tensor embedding)	Forcibly sets an embedding tensor.
mklab.JGNN.core.Tensor	getEmbedding()	Retrieves the embedding of the node.
java.util.LinkedList<mklab.JGNN.core.Tensor>	getEmbeddingHistory()
mklab.JGNN.core.Tensor	getNeighborAggregation()	An aggregation of the node's neighborhood embeddings in the social graph.
GNNNodeData	setLearningRate(double learningRate)	Sets the learning rate (default is 1) of the updateEmbedding(Tensor) operation.
void	setNeighborAggregation(mklab.JGNN.core.Tensor neighborAggregation)	Sets a neighbor aggregation that can be retrieved with getNeighborAggregation().
void	setRegularization(mklab.JGNN.core.Tensor regularization)	Sets the regularization (default is a zero vector) of the updateEmbedding(Tensor) operation.
GNNNodeData	setRegularizationWeight(double regularizationWeight)	Sets the regularization weight (default is 0.1) of the updateEmbedding(Tensor) operation.
void	updateEmbedding(mklab.JGNN.core.Tensor derivative)	Performs the operation embedding += (embedding-regularization)*learningRate*regularizationWeight-derivative*learningRate that is a regularized gradient descent over a computed derivative, where the area of regularization is constrained towards the point set by setRegularization(Tensor).

Methods inherited from class java.lang.Object

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

Constructor Detail

GNNNodeData

public GNNNodeData()

Method Detail

addEmbeddingToHistory

public void addEmbeddingToHistory()

addRegularizationToHistory

public void addRegularizationToHistory()

getEmbeddingHistory

public java.util.LinkedList<mklab.JGNN.core.Tensor> getEmbeddingHistory()

getEmbedding

public mklab.JGNN.core.Tensor getEmbedding()

Retrieves the embedding of the node.

Returns:

A Tensor holding the embedding representation.

forceSetEmbedding

public void forceSetEmbedding(mklab.JGNN.core.Tensor embedding)

Forcibly sets an embedding tensor. Normally, embeddings are updated through updateEmbedding(Tensor).

Parameters:

embedding - The embedding Tensor

setNeighborAggregation

public void setNeighborAggregation(mklab.JGNN.core.Tensor neighborAggregation)

Sets a neighbor aggregation that can be retrieved with getNeighborAggregation(). This aggregation is computed by other devices and this function is called when receiving it as part of the shared parameters. These operations are automatically performed by GNNMiner.newInteractionParameters(eu.h2020.helios_social.core.contextualegonetwork.Interaction, eu.h2020.helios_social.modules.socialgraphmining.SocialGraphMiner.SocialGraphMinerParameters, eu.h2020.helios_social.modules.socialgraphmining.SocialGraphMiner.InteractionType)

Parameters:

neighborAggregation - The received Tensor of neighbor aggregation

getNeighborAggregation

public mklab.JGNN.core.Tensor getNeighborAggregation()

An aggregation of the node's neighborhood embeddings in the social graph.

Returns:

A tensor holding the aggregated neighborhood embeddings

See Also:

setNeighborAggregation(Tensor)

setRegularization

public void setRegularization(mklab.JGNN.core.Tensor regularization)

Sets the regularization (default is a zero vector) of the updateEmbedding(Tensor) operation. Setting this to a value other than a zero tensor helps influence the trained embedding to a latent space around the given value. Hence, this can be used to influence the embeddings of nodes in the contextual ego network given their embeddings in other devices.

Parameters:

regularization - The given regularization tensor.

See Also:

setRegularizationWeight(double)

setLearningRate

public GNNNodeData setLearningRate(double learningRate)

Sets the learning rate (default is 1) of the updateEmbedding(Tensor) operation.

Parameters:

learningRate - The given regularization weight.

Returns:

this GNNNodeData instance.

setRegularizationWeight

public GNNNodeData setRegularizationWeight(double regularizationWeight)

Sets the regularization weight (default is 0.1) of the updateEmbedding(Tensor) operation. This helps influence the trained embedding to a latent space around the given value

Parameters:

regularizationWeight - The given regularization weight.

Returns:

this GNNNodeData instance.

See Also:

setRegularization(Tensor)

updateEmbedding

public void updateEmbedding(mklab.JGNN.core.Tensor derivative)

Performs the operation embedding += (embedding-regularization)*learningRate*regularizationWeight-derivative*learningRate that is a regularized gradient descent over a computed derivative, where the area of regularization is constrained towards the point set by setRegularization(Tensor).

Parameters:

derivative - The derivative of the embedding.

See Also:

getEmbedding(), setLearningRate(double), setRegularizationWeight(double), setRegularization(Tensor)