netkit.classifiers.active
Class GraphCentralityLabeling

java.lang.Object
  netkit.classifiers.active.PickLabelStrategyImp
      netkit.classifiers.active.GraphCentralityLabeling

All Implemented Interfaces:

PickLabelStrategy, Configurable

public class GraphCentralityLabeling
extends PickLabelStrategyImp

Graph Centrality Labeling for Active Learning iteratively picks central nodes in a graph that are in clusters that have no known labels. It uses a modularity based clustering algorithm to iteratively cluster a graph into smaller and smaller clusters until it finds a cluster that has no currently known labels. It then picks the most central node in that cluster.

Various centrality metric strategies are possible (using the metric configuration variable):

• metric --- name of the class (full class name or stem if in the netkit.classifiers.active.graphfunctions package)
• useClustering=[true|false] -- cluster graph and find central nodes in each cluster?

If clustering is turned on, then various cluster-ranking functions are available (after using the centrality metric to identify the node in the cluster). The cluster rank function is specified using the clusterrank configuration variable. By default, the rank is the same as the centrality metric.

• clusterRank --- name of the class (full class name or stem if in the netkit.classifiers.active.graphfunctions package)
• nodesPerCluster --- keep the topk list in each cluster and return those as well (first topk from first cluster, then iteratively through the next clusters). default topk=1

Author:

sofmac

See Also:

netkit.graph.ModularityClusterer

Nested Class Summary

Nested classes/interfaces inherited from interface netkit.classifiers.active.PickLabelStrategy

PickLabelStrategy.LabelNode

Field Summary

Fields inherited from class netkit.classifiers.active.PickLabelStrategyImp

iteration, logger

Constructor Summary

GraphCentralityLabeling()


Method Summary

void	configure(Configuration config)
Configuration	getDefaultConfiguration()
java.lang.String	getDescription()
Harmonic	getHarmonicFunction()
static GraphCentralityLabeling	getInstance(java.lang.String metric, java.lang.String rank, double nodesPerCluster)
Classification	getLabels()
GraphMetrics	getMetrics()
java.lang.String	getName()
double	getRank(DataSplit split, Estimate predictions, Node node) Get the rank of the given node if the strategy were to pick the node.
java.lang.String	getShortName()
void	initialize(NetworkLearner nl, DataSplit split) Initialize the label strategy by providing a reference to the NetworkLeaner object that calls the strategy, thereby giving it access to all information it is likely to need.
PickLabelStrategy.LabelNode[]	peek(DataSplit split, Estimate predictions, int maxPicks) Get the list of nodes to get labels for...
protected PickLabelStrategy.LabelNode[]	pickNodes(Estimate predictions, int maxPicks) Picks the next nodes as the ones with the highest closeness centrality (normalized by cluster size) in a cluster that has no known labels.
protected void	printStat(java.lang.String title)

Methods inherited from class netkit.classifiers.active.PickLabelStrategyImp

getAverageRank, getAverageRank, getIterationNum, getNetworkLearner, getNodesToLabel, getSplit, toString

Methods inherited from class java.lang.Object

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

Constructor Detail

GraphCentralityLabeling

public GraphCentralityLabeling()

Method Detail

getInstance

public static GraphCentralityLabeling getInstance(java.lang.String metric,
                                                  java.lang.String rank,
                                                  double nodesPerCluster)

getDefaultConfiguration

public Configuration getDefaultConfiguration()

Specified by:

getDefaultConfiguration in interface Configurable

Overrides:

getDefaultConfiguration in class PickLabelStrategyImp

configure

public void configure(Configuration config)

Specified by:

configure in interface Configurable

Overrides:

configure in class PickLabelStrategyImp

getMetrics

public GraphMetrics getMetrics()

getLabels

public Classification getLabels()

getHarmonicFunction

public Harmonic getHarmonicFunction()

initialize

public void initialize(NetworkLearner nl,
                       DataSplit split)

Description copied from interface: PickLabelStrategy

Initialize the label strategy by providing a reference to the NetworkLeaner object that calls the strategy, thereby giving it access to all information it is likely to need. This should re-initialize the strategy if it has any cached values from prior calls.

Specified by:

initialize in interface PickLabelStrategy

Overrides:

initialize in class PickLabelStrategyImp

Parameters:

nl - The NetworkLearner object that will be calling this strategy.

split - The initial train/test split that this will be used on

getDescription

public java.lang.String getDescription()

getName

public java.lang.String getName()

getShortName

public java.lang.String getShortName()

printStat

protected void printStat(java.lang.String title)

getRank

public double getRank(DataSplit split,
                      Estimate predictions,
                      Node node)

Description copied from interface: PickLabelStrategy

Get the rank of the given node if the strategy were to pick the node. If more than one node has the same score, then average their ranks (hence a double is returned)

Specified by:

getRank in interface PickLabelStrategy

Overrides:

getRank in class PickLabelStrategyImp

Parameters:

split - Current datasplit

predictions - Current predictions of the classifier

node - the node whose rank is requested

Returns:

the rank of the given node or Double.NaN if the node is not in the rankings. If more than one node has the same score, then average their ranks (hence a double is returned)

peek

public PickLabelStrategy.LabelNode[] peek(DataSplit split,
                                          Estimate predictions,
                                          int maxPicks)

Description copied from interface: PickLabelStrategy

Get the list of nodes to get labels for... without changing the internal state of the active labeler.

Specified by:

peek in interface PickLabelStrategy

Overrides:

peek in class PickLabelStrategyImp

Parameters:

split - Current datasplit

predictions - Current predictions of the classifier

Returns:

An array of Node objects that should receive labels. null is returned if done.

pickNodes

protected PickLabelStrategy.LabelNode[] pickNodes(Estimate predictions,
                                                  int maxPicks)

Picks the next nodes as the ones with the highest closeness centrality (normalized by cluster size) in a cluster that has no known labels. As it chooses new labels, it will split those clusters into two clusters: one with a labeled example and one with no labeled examples.

Specified by:

pickNodes in class PickLabelStrategyImp

maxPicks - how many nodes should it pick at maximum (this iteration)

Returns:

An array of Node objects that should receive labels. null is returned if done.