mulan.classifier.neural
Class BPMLLAlgorithm

java.lang.Object
  mulan.classifier.neural.BPMLLAlgorithm

public class BPMLLAlgorithm
extends Object

The implementation of Back-Propagation Multi-Label Learning (BPMLL) algorithm for neural networks. The algorithm uses weights decay regularization to avoid overfitting.

For more information see:
Zhang, M.L., Zhou, Z.H.: Multi-label neural networks with applications to functional genomics and text categorization. IEEE Transactions on Knowledge and Data Engineering 18 (2006) 1338-1351

Version:

2012.02.27

Author:

Jozef Vilcek

See Also:

NeuralNet

Constructor Summary

BPMLLAlgorithm(NeuralNet neuralNet, double weightsDecayCost)
Creates a BPMLLAlgorithm instance.

Method Summary

NeuralNet	getNetwork() Returns the neural network which is learned/updated by the algorithm.
double	getNetworkError(double[] inputPattern, double[] expectedLabels) Returns the error of the neural network for given input.
double	getWeightsDecayCost() Returns the value of weights decay cost term used for regularization.
double	learn(double[] inputPattern, double[] expectedLabels, double learningRate) Performs one learning step with given input pattern and expected output values.

Methods inherited from class java.lang.Object

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

Constructor Detail

BPMLLAlgorithm

public BPMLLAlgorithm(NeuralNet neuralNet,
                      double weightsDecayCost)

Creates a BPMLLAlgorithm instance.

Parameters:

neuralNet - the neural network model to learn

weightsDecayCost - the weights decay cost term used for regularization. The value must be greater than 0 and no more than 1.

Method Detail

getNetwork

public NeuralNet getNetwork()

Returns the neural network which is learned/updated by the algorithm.

Returns:

the neural network

getWeightsDecayCost

public double getWeightsDecayCost()

Returns the value of weights decay cost term used for regularization.

Returns:

the weights decay cost term

learn

public double learn(double[] inputPattern,
                    double[] expectedLabels,
                    double learningRate)

Performs one learning step with given input pattern and expected output values. The function outputs the error for passed input pattern.
The input is ignored by the algorithm (can not process) if the input example has assigned either all or non of the labels. In this case, the function returns Double.NaN.

Parameters:

inputPattern - the input pattern for the network

expectedLabels - the ideal, expected values the network should output as a response for the given input. If the i-th label class belongs to the input pattern instance, then i-th value is +1, otherwise the value is -1.

learningRate - the learning rate used to update the neural network weights

Returns:

the error of the network response for the passed input or Double.NaN if the passed input can not be processed.

getNetworkError

public double getNetworkError(double[] inputPattern,
                              double[] expectedLabels)

Returns the error of the neural network for given input. This is value of error function computed from network output value for given input and expected, ideal output for given input.
The input is ignored by the algorithm (can not process), if the input example has assigned either all or non of labels. In this case, the function returns Double.NaN.

Parameters:

inputPattern - the input pattern to be processed

expectedLabels - the ideal, expected values the network should output as a response for the given input. If the ith label class belongs to the input pattern instance, then ith value is +1, otherwise the value is -1.

Returns:

the error of the network response for the passed input or Double.NaN if the passed input can not be processed