Machine Learning: WEKA Tutorial Kecerdasan Buatan Setio Basuki, ST., MT

(1)

Machine Learning:

WEKA Tutorial

(2)

Practicing WEKA

• What is WEKA ?

• Formatting the data into ARFF

• Klasifkasi

– Tahapan membangun classifer

– Contoh kasus : Klasifkasi bunga iris

– Tahapan membangun classifer

– Merangkum hasil eksperimen k-Nearest

Neighbor Classifer

– Eksperimen memakai classifer yang lain (JST,

SVM)

(3)

3

What is WEKA ?

• Machine learning/data mining software written

in Java (distributed under the GNU Public

License)

• Used for research, education, and applications

• Complements “Data Mining” by Witten & Frank

• Main features:

– Comprehensive set of data pre-processing tools,

learning algorithms and evaluation methods

– Graphical user interfaces (incl. data visualization)

– Environment for comparing learning algorithms

• Weka versions

– WEKA 3.4: “book version” compatible with

description in data mining book

(4)

Formatting Data into ARFF

@relation iris

@attribute sepallength real

@attribute sepalwidth real

@attribute petallength real

@attribute petalwidth real

(5)

5

Practicing WEKA

• What is WEKA ?

• Formatting the data into ARFF

• _Klasifkasi

– Tahapan membangun classifer

– Contoh kasus : Klasifkasi bunga iris

– Tahapan membangun classifer

– Merangkum hasil eksperimen k-Nearest

Neighbor Classifer

– Eksperimen memakai classifer yang lain (JST,

SVM)

– Classifcation of cancers based on gene

expression

– Parkinson Disease Detection

(6)

Tahapan membangun Classifer

1. Tentukan manakah informasi yang merupakan

(a) attribute/feature

(b) class

(c) training & testing set

(d) skenario pengukuran akurasi

2. Tentukan kombinasi parameter model, dan

(7)

7

Contoh Kasus : Klasifkasi bunga iris

• Data set yang paling terkenal

• Author: R.A. Fisher

• Terdiri dari 3 kelas,

masing-masing memiliki 50 samples

(instances)

• Attribute information:

– Sepal (kelopak) length in cm

– sepal width in cm

– Petal (mahkota) length in cm

– petal width in cm

– class: (1) Iris Setosa (2) Iris

Versicolour (3)Iris Virginica

(8)

(9)

9

Tahapan membangun Classifer

1. Tentukan manakah informasi yang merupakan

(a) attribute/feature : sepal length (panjang

(b) class:

iris setosa

iris versicolor

iris virginica

(c) training & testing set

(10)

(11)

11

(12)

(13)

13

(14)

(15)

15

IB1 : 1-Nearest Neighbor

Classifer)

(16)

(17)

17

SMO singkatan dari

(18)

(19)

19

Misalnya kita pilih

(20)

Selanjutnya pilihlah skenario

Pengukuran akurasi. Dari 4

Options yang diberikan, pilihlah

“Supplied test set” dan klik

Button “Set” untuk memiilih

(21)

21

Tahapan membangun Classifer

Iris-training.arf

1. Naïve Bayes

2. K-Nearest Neighbor Classifer

(lazy



iBk)

3. Artifcial Neural Network

(function



multilayer

perceptron)

(22)

Apakah yang dimaksud “mengukur akurasi”

• Testing set “iris-testing.arf” dilengkapi dengan

(23)

23

Berbagai cara pengukuran akurasi

• “Using training set” : memakai seluruh data sebagai training set, sekaligus

testing set. Akurasi akan sangat tinggi, tetapi tidak memberikan estimasi

akurasi yang sebenarnya terhadap data yang lain (yang tidak dipakai untuk

training)

• Hold Out Method : Memakai sebagian data sebagai training set, dan sisanya sebagai testing set. Metode yang lazim dipakai, asal jumlah sampel cukup

(24)

Berbagai cara pengukuran akurasi

• Cross Validation Method

( fold = 5 atau 10 ) : teknik

estimasi akurasi yang

dipakai, jika jumlah sampel

terbatas. Salah satu bentuk

khusus CV adalah

Leave-one-out Cross Validation

(LOOCV) : dipakai jka

(25)

25

Ilustrasi Cross Validation (k=5)

1. Data terdiri dari 100 instances (samples), dibagi ke dalam 5 blok dengan jumlah sampel yang sama. Nama blok : A, B, C, D dan E, masing-masing terdiri dari 20 instances

2. Kualitas kombinasi parameter tertentu diuji dengan cara sbb.

step 1:

training memakai A,B,C,D testing memakai E

akurasi a

step 2:

training memakai A,B,C,E testing memakai D

akurasi b

step 3:

training memakai A,B, D,E testing memakai C

akurasi c

step 4:

training memakai A, C,D,E testing memakai B

akurasi d

step 5:

training memakai B,C,D,E testing memakai A

akurasi e

3. Rata-rata akurasi : (a+b+c+d+e)/5 mencerminkan kualitas parameter yang dipilih

(26)

(27)

27

(28)

Klik button “Start”

(29)

• Baris pertama “25 0 0” menunjukkan bahwa ada (25+0+0)

instances class Iris-setosa di dalam fle iris-testing.arf dan

semua benar diklasifkasikan sebagai Iris setosa

• Baris kedua “0 24 1” menunjukkan bahwa ada (0+24+1)

instances class Iris-versicolor di dalam fle iris-testing.arf dan

1 salah diklasifkasikan sebagai Iris-virginica

• Baris ketiga “0 2 24” menunjukkan bahwa ada (0+2+23)

(30)

(31)

Inst# : nomer urut data pada fle

“iris-testing.arf”

actual : class yang sebenarnya

predicted: class yang diprediksi

Error: jika ada misclassifcation, akan diberikan

tanda “+”

(32)

Merangkum hasil eksperimen

No.

K

Correct Classification Rate

Iris setosa

Iris versicolor

Iris virginica

Total

(33)

33

Eksperimen memakai Neural Network

• Untuk eksperimen memakai

neural network, caranya sama

dengan k-Nearest Neighbor

Classifer.

• Parameter yang dituning

meliputi antara lain:

– hiddenLayers: banyaknya

neuron pada hidden layer.

Default “a” : rata-rata jumlah

neuron pada input & output

layer

– LearningRate : biasanya nilai

kecil (0.1, 0.01, 0.2, 0.3 dsb)

– Momentum: biasanya nilai

besar (0.6, 0.9 dsb)

– trainingTime: maksimum

(34)

(35)

35

Eksperimen memakai SVM

C: complexity

parameter (biasanya

mengambil nilai

besar. 100, 1000

dst)

(36)

(37)

37

Classifcation of cancers based on

gene expression

• Biological reference:

Classifcation and diagnostic prediction of cancers using

gene expression profling and artifcial neural networks,

J. Khan, et al., Nature Medicine 7, pp.673-679, 2001

(http://www.thep.lu.se/~carsten/pubs/lu_tp_01_06.pdf )

• Data is available from

http://research.nhgri.nih.gov/microarray/Supplement/

• Small Round Blue Cell Tumors (SRBCT) has two class:

– Ewing Family of Tumors (EWS)

– NB: Neuroblastoma

– BL: Burkitt lymphomas

– RMS: Rhabdomyosarcoma : RMS

• Characteristic of the data

– Training samples

:

63 (EWS:23 BL:8 NB:12

RMS:20)

(38)

Experiment using k-Nearest Neighbor Classifer

• Training and testing set are given as

separated arf fle

• Use training set to build a classifer: k-Nearest

Neighbor (k=1)

• Evaluate its performance on the testing set.

• Change the value of k into 3,5,7 and 9 and

repeat step 1 to 3 for each value.

Experiment using Artifcial Neural Network

• Do the same experiment using Multilayer

(39)

39

Parkinson Disease Detection

Max Little (Oxford University) recorded speech signals and measured the biomedical voice from 31 people, 23 with Parkinson Disease (PD). In the dataset which will be distributed during fnal examination,

each column in the table is a particular voice measure, and each row corresponds one of 195 voice recording from these individuals

("name" column). The main aim of the data is to discriminate healthy people from those with PD, according to "status" column which is set to 0 for healthy and 1 for PD. There are around six recordings per patient, making a total of 195 instances. (Ref. 'Exploiting Nonlinear Recurrence and Fractal Scaling Properties for Voice Disorder

Detection', Little MA, McSharry PE, Roberts SJ, Costello DAE, Moroz IM. BioMedical Engineering OnLine 2007, 6:23, 26 June 2007).

Experiment using k-Nearest Neighbor Classifer

Conduct classifcation experiments using k-Nearest Neighbor

(40)

Parkinson Disease Detection

Max Little (Oxford University) recorded speech signals and measured the biomedical voice from 31 people, 23 with Parkinson Disease (PD). In the dataset which will be distributed during fnal examination,

each column in the table is a particular voice measure, and each row corresponds one of 195 voice recording from these individuals

("name" column). The main aim of the data is to discriminate healthy people from those with PD, according to "status" column which is set to 0 for healthy and 1 for PD. There are around six recordings per patient, making a total of 195 instances. (Ref. 'Exploiting Nonlinear Recurrence and Fractal Scaling Properties for Voice Disorder

Detection', Little MA, McSharry PE, Roberts SJ, Costello DAE, Moroz IM. BioMedical Engineering OnLine 2007, 6:23, 26 June 2007).