23.05.2014 goal of the project classification of …...23.05.2014 1 classification of haberman...

9
23.05.2014 1 Classification of Haberman Survival dataset EECS 589 Artificial Neural networks By Pınar Uskaner 21.05.2014 Goal of the project Train the neural network to predict the class (dead or alive) whether a patient survived after breast cancer surgery, when it is given other attributes as input. As a result of this survey, we want to find optimal neural network model in order to generalize Haberman’s survival problem by using neural nerons. Inputs Outputs Age of patient at time of operation (numerical) Year of operation (year - 1900, numerical) Number of positive axillary nodes detected (numerical) 1= the patient survived 5 years or longer 2 = the patient died within 5 year Methods: The type of neural network : Multi layer perceptron Training algorithm: Resilient backpropagation Normalization is used to be able to train data faster Sigmoid transfer function (logsig) is used to constrain the outputs of a network (such as between 0 and 1).

Upload: others

Post on 28-May-2020

14 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: 23.05.2014 Goal of the project Classification of …...23.05.2014 1 Classification of Haberman Survival dataset EECS 589 Artificial Neural networks By Pınar Uskaner network model

23.05.2014

1

Classification of Haberman Survival dataset EECS 589 Artificial Neural networks By Pınar Uskaner 21.05.2014

Goal of the project

• Train the neural network to predict the class (dead or alive) whether a patient survived after breast cancer surgery, when it is given other attributes as input.

• As a result of this survey, we want to find optimal neural network model in order to generalize Haberman’s survival problem by using neural nerons.

Inputs Outputs

•Age of patient at time of operation (numerical)

• Year of operation (year - 1900, numerical)

•Number of positive axillary nodes detected (numerical)

• 1= the patient survived 5 years or longer

• 2 = the patient died within 5 year

Methods:

• The type of neural network : Multi layer perceptron

• Training algorithm: Resilient backpropagation

•Normalization is used to be able to train data faster

• Sigmoid transfer function (logsig) is used to constrain the outputs of a network (such as between 0 and 1).

Page 2: 23.05.2014 Goal of the project Classification of …...23.05.2014 1 Classification of Haberman Survival dataset EECS 589 Artificial Neural networks By Pınar Uskaner network model

23.05.2014

2

Procedure of training a neural network

• Preparation the data set

•Create a training set

•Create a neural network

• Train the network

• Test the network to make sure that it is trained properly

Normalization

• The standart MaxMin formula:

B=((A-minValue of A)/(maxValue of A-minValue of A))*(D-C)+C B is the standardized value A the given value D and C determine the range in which we want our value to be. In this case, D= 0 and C=1

Training

• Training is performed by giving the neural network a set of sample data along with the anticipated outputs from each of these samples

• Training proceeds, the neural network is taken through a number of iterations, until the output of the neural network matches the anticipated output, with a reasonably small rate of error.

Creating a neural network

•Decision of the number of hidden layer neurons

•Max error

• Set learning parameters • Learning rate:

• Higher learning rate speeds up the process of learning

• Momentum: • Decreasing momentum value causes growing iteration numbers

Rules for number of hidden neurons

• The number of hidden neurons should be between the size of the input layer and the size of the output layer

• The number of hidden neurons should be 2/3 the size of the input layer, plus the size of the output layer

• The number of hidden neurons should be less than

twice the size of the input layer

Training dataset Multi layer perceptron parameters:

• input neurons: 3 •hidden neurons: 2 •output neurons: 1 • Transfer function: sigmoid • learning rule: Backpropagation with Momentum •Max Error:0.02

Page 3: 23.05.2014 Goal of the project Classification of …...23.05.2014 1 Classification of Haberman Survival dataset EECS 589 Artificial Neural networks By Pınar Uskaner network model

23.05.2014

3

Cont.

Learning parameters:

• Learning Rate:0.3 •Momentum:0.4

Neural network tool

Mean squared errors Mean squared error

The default performance function for feedforward networks

is mean square error mse—the average squared error between the network outputs a and the target outputs t.

Matlab

Total error (0.3045) is quitely larger than max error value. So we need to improve the neural network.

Training dataset Multi layer perceptron parameters:

• input neurons: 3 •hidden neurons: 3 •output neurons: 2 • Transfer function: sigmoid • learning rule: Backpropagation with Momentum •Max Error:0.01

Page 4: 23.05.2014 Goal of the project Classification of …...23.05.2014 1 Classification of Haberman Survival dataset EECS 589 Artificial Neural networks By Pınar Uskaner network model

23.05.2014

4

Cont.

Learning parameters:

• Learning Rate:0.2 •Momentum:0.7

Neural network tool

Matlab

Total error (0.0042) is quitely smaller than max error value. So we can test the network.

Testing

• Total mean square value (0.1701) is close to zero. We will choose random 5 input values from dataset and test them.

İNPUTS TARGETS OUTPUTS

0,320754716981132 0 0,0576923076923077 0 0.2419

0,320754716981132 0,181818181818182 0,0769230769230769 0 0.2025

0,320754716981132 0,909090909090909 0,0769230769230769 0 0.2523

0,320754716981132 0,727272727272727 0,230769230769231 0 0.4212

0,339622641509434 0 0,211538461538462 1 0.5778

Cont.

• The results above table showed that the network predicted all the output values as correct values.

• This network may be a solution to the problem

Training the same network with different momentum value

•Learning parameters: Learning Rate:0.3 Momentum:0.7

Page 5: 23.05.2014 Goal of the project Classification of …...23.05.2014 1 Classification of Haberman Survival dataset EECS 589 Artificial Neural networks By Pınar Uskaner network model

23.05.2014

5

Matlab

Sum of the error (0.2623) is larger than max error. Changing the momentum value doesn’t help much.

Training the same network with different learning rate value

•Learning parameters: Learning Rate:0.5 Momentum:0.7

Matlab

Total error is 0.0047. It is smaller than max error. We can test it.

Testing

• Total mean square value (0.1923) is close to zero. We will choose random 5 input values from dataset and test them.

INPUTS TARGETS OUTPUTS 0,2830 0,9090 0 0 0.2970

0,2830 0,8181 0,0192 0 0.2505

0,3018 0 0,0384 1 0.6909

0,3018 1 0,0576 1 0.7563

0,3018 0,3636 0,0961 1 0.5987

Results

• Total mean square error value is 0.1923 which is close to

zero. Hovewer, we test randomly chosen 5 input values.

The results above table showed that the network

predicted all the output values as correct values

Training dataset Multi layer perceptron parameters:

• input neurons: 3 •hidden neurons: 4 •output neurons: 2 • Transfer function: sigmoid • learning rule: Backpropagation with Momentum •Max Error:0.01

Page 6: 23.05.2014 Goal of the project Classification of …...23.05.2014 1 Classification of Haberman Survival dataset EECS 589 Artificial Neural networks By Pınar Uskaner network model

23.05.2014

6

Cont.

Learning parameters:

• Learning Rate:0.5 •Momentum:0.7

Matlab

Total error is smaller than max error. So we can test the network

Testing

• Total mean square value (0.2564) is close to zero. We have better results. So we do not need to test randomly chosen input values from the dataset.

Training the same network with different momentum value

•Learning parameters: Learning Rate:0.5 Momentum:0.6

Page 7: 23.05.2014 Goal of the project Classification of …...23.05.2014 1 Classification of Haberman Survival dataset EECS 589 Artificial Neural networks By Pınar Uskaner network model

23.05.2014

7

Training the same network with different momentum value

•Learning parameters: Learning Rate:0.5 Momentum:0.4

Matlab

Total error =0.1980

It is to big. Decreasing momentum value causes growing iteration number.

Training dataset Multi layer perceptron parameters:

• input neurons: 3 •hidden neurons: 6 •output neurons: 1 • Transfer function: sigmoid • learning rule: Backpropagation with Momentum •Max Error:0.01

Cont.

Learning parameters:

• Learning Rate:0.2 •Momentum:0.7

Page 8: 23.05.2014 Goal of the project Classification of …...23.05.2014 1 Classification of Haberman Survival dataset EECS 589 Artificial Neural networks By Pınar Uskaner network model

23.05.2014

8

Matlab

Total error is smaller than max error. So we can test the network

Testing

• Total mean square value (0.1795) is close to zero. We will choose random 5 input values from dataset and test them.

INPUTS TARGETS OUTPUTS

0,4905 0,7272 0,0384 0 0.1449 0,4905 0,7272 0,0192 0 0.1451 0,4905 0,8181 0 0 0.1422 0,4905 0,1818 0 0 0.1654 0,5094 0,2727 0,0961 1 0.5372

Cont.

• The network guessed all of them right. We can conclude that this network has a good ability of generalization

The best two optimal neural network for the survival dataset:

• MLP with 3 hidden neurons:

• Learning rate:0.2

• Momentum:0.7

• Max Error=0.01

• Total MSE=0.1701

• MLP with 6 hidden neurons:

• Learning rate:0.2

• Momentum:0.7

• Max Error=0.01

• Total MSE=0.1795

Conclusions

• It is shown that how neural networks is used for classification problems.

• Several architecture of NNs were worked and was decided which one is the best solution to the Haberman Survival problem

• The aim is to learn machine with the least error and get the closest output to the desired output.

• Different solutions which were tested in this experiment have shown that the success of a neural network is very sensitive to parameters choice in the training process.

Page 9: 23.05.2014 Goal of the project Classification of …...23.05.2014 1 Classification of Haberman Survival dataset EECS 589 Artificial Neural networks By Pınar Uskaner network model

23.05.2014

9

Cont.

• The learning rate must not be too high, because higher learning rate speeds up the process of learning. And also, the maximum error must not be too low.

• Total mean square error does not reflect directly the success of a network.

•Deciding the number of hidden neurons is important, because it affects performance of a neural network.

References

• Dursun Delen, Glenn Walker, Amit Kadami, Predicting breast cancer survivability: a comparison of three data mining methods Artificial intelligence in medicine 1 June 2005 (volume 34 issue 2 Pages 113-127 DOI: 10.1016/j.artmed.2004.07.002)

• Hussein A. Abbass, An Evolutionary Artificial Neural Networks Approach for Breast Cancer Diagnosis, School of Computer Science, University of New South Wales, Australian Defence Force Academy Campus

References

• Filippo Amato, Alberto López, Eladia María Peña-Méndez, Petr Vaňhara, Aleš Hampl,Josef Havel, Artificial Neural Networksn in Medical Diagnosis, Journal of Applied Medicine

• Yann Le Cun, Leon Botou, Genevieve B. Orr, Klaus Robert Muller, Efficient Backpropagation, Wilamette University, USA

• Mark Hudson Beale, Martin T. Hagan, Howard B. Demuth , Neural Network Toolbox Guide