netkit.classifiers
Class NetworkLearner

java.lang.Object
  netkit.classifiers.NetworkLearner

public class NetworkLearner
extends java.lang.Object

Field Summary

java.util.logging.Logger

logger

Constructor Summary

NetworkLearner(Classifier lc, NetworkClassifier nc, InferenceMethod ic, boolean applyCMN)

Method Summary

Graph	getGraph()
InferenceMethod	getInferenceMethod()
Classifier	getLocalClassifier()
NetworkClassifier	getNetworkClassifier()
DataSplit	getSplit()
boolean	inTrainingSet(Node n)
Estimate	runActiveLearner(PickLabelStrategy ps, DataSplit split) See fully parameterized method for details.
Estimate	runActiveLearner(PickLabelStrategy ps, DataSplit split, int picksPerIteration, int maxPicks, boolean learnWithTruth, int depth) Run active learning using the given parameters.
Estimate	runInference(DataSplit split)
Estimate	runInference(DataSplit split, boolean showItAcc, boolean learnWithTruth, int depth)
Estimate	runLeaveOneOut(DataSplit split)
Estimate	runLeaveOneOut(DataSplit split, boolean learnWithTruth, int depth)
void	saveIterationPredictions(java.lang.String stem, PrintEstimateWriter pe, boolean append, Node[] eval, java.lang.String header)
void	saveIterationsInPajek(java.lang.String pajekFile)

Methods inherited from class java.lang.Object

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

Field Detail

logger

public final java.util.logging.Logger logger

Constructor Detail

NetworkLearner

public NetworkLearner(Classifier lc,
                      NetworkClassifier nc,
                      InferenceMethod ic,
                      boolean applyCMN)

Method Detail

getNetworkClassifier

public NetworkClassifier getNetworkClassifier()

getLocalClassifier

public Classifier getLocalClassifier()

getInferenceMethod

public InferenceMethod getInferenceMethod()

getGraph

public Graph getGraph()

getSplit

public DataSplit getSplit()

inTrainingSet

public boolean inTrainingSet(Node n)

runActiveLearner

public Estimate runActiveLearner(PickLabelStrategy ps,
                                 DataSplit split)

See fully parameterized method for details. This method calls the more fully qualified method, with the following parameters:

runActiveLearner(ps,split,seedSize,true)

It computes seedSize to be the number of classes possible, thereby in effect setting the initial seeds to be one instance of each class randomly chosen from the training set from the given DataSplit.

Parameters:

ps - Strategy for picking labels

split - Split from which to pick seeds + test set

Returns:

Final estimates after running active learning

See Also:

#runActiveLearner(PickLabelStrategy, DataSplit, boolean, int, boolean)

runActiveLearner

public Estimate runActiveLearner(PickLabelStrategy ps,
                                 DataSplit split,
                                 int picksPerIteration,
                                 int maxPicks,
                                 boolean learnWithTruth,
                                 int depth)

Run active learning using the given parameters.

The methodology works as follows:

1. Initialize the strategy object with a reference to this object
2. Repeat until labelSet is as large as maximum size given or the strategy does not pick more nodes.
   1. Learn a model using the labelSet. Use complete truth (learnWithTruth) or truth up to depth from nodes in the labelSet
   2. Use learned model to predict labels from all nodes that are not in the labelSet.
   3. Use strategy to pick nodes that should be labeled next. If pickFromTrainingOnly is true, then only pick from nodes that are in the DataSplit training set, otherwise allow the strategy to pick any node in the graph that are not part of the DataSplit test set.
   4. Get true labels from the picked nodes and add these to labelSet.
   Return the final predictions for the test set only

Parameters:

ps - Strategy for picking labels

split - Split from which to pick seeds + test set

picksPerIteration - How many instances should be picked each iteration

maxPicks - How many instances should be picked total

learnWithTruth - Use complete truth when learning classifiers

depth - Get true labels to this depth when learning a relational classifier

Returns:

Final estimates after running active learning

runLeaveOneOut

public Estimate runLeaveOneOut(DataSplit split)

runLeaveOneOut

public Estimate runLeaveOneOut(DataSplit split,
                               boolean learnWithTruth,
                               int depth)

saveIterationPredictions

public void saveIterationPredictions(java.lang.String stem,
                                     PrintEstimateWriter pe,
                                     boolean append,
                                     Node[] eval,
                                     java.lang.String header)

saveIterationsInPajek

public void saveIterationsInPajek(java.lang.String pajekFile)

runInference

public Estimate runInference(DataSplit split)

runInference

public Estimate runInference(DataSplit split,
                             boolean showItAcc,
                             boolean learnWithTruth,
                             int depth)