1

Machine Learning Methodsfor

Text / Web Data Mining

Byoung-Tak ZhangSchool of Computer Science and Engineering

Seoul National UniversityE-mail: btzhang@cse.snu.ac.kr

This material is available at http://scai.snu.ac.kr./~btzhang/

2

Overview

Introduction4Web Information Retrieval 4Machine Learning (ML)4ML Methods for Text/Web Data Mining

Text/Web Data Analysis4Text Mining Using Helmholtz Machines4Web Mining Using Bayesian Networks

Summary 4Current and Future Work

2

3

Web Information RetrievalText Data

Classification System

Information Filtering System

questionuser profile

feedback

answer

DB

LocationDate

DB Record

DB Template Filling& InformationExtraction System

Information FilteringInformation Extraction

filtered data

Text Classification

Preprocessing and Indexing

4

Machine Learning

Supervised Learning4Estimate an unknown mapping from known input-

output pairs4Learn fw from training set D={(x,y)} s.t.4Classification: y is discrete4Regression: y is continuous

Unsupervised Learning4Only input values are provided4Learn fw from D={(x)} s.t.4Density Estimation4Compression, Clustering

)()( xxw fyf ==

xxw =)(f

3

5

Machine Learning MethodsNeural Networks4Multilayer Perceptrons (MLPs)4Self-Organizing Maps (SOMs)4Support Vector Machines (SVMs)

Probabilistic Models4Bayesian Networks (BNs)4Helmholtz Machines (HMs)4Latent Variable Models (LVMs)

Other Machine Learning Methods4Evolutionary Algorithms (EAs)4Reinforcement Learning (RL)4Boosting Algorithms4Decision Trees (DTs)

6

ML for Text/Web Data Mining

Bayesian Networks for Text ClassificationHelmholtz Machines for Text Clustering/CategorizationLatent Variable Models for Topic Word ExtractionBoosted Learning for TREC Filtering TaskEvolutionary Learning for Web Document RetrievalReinforcement Learning for Web Filtering AgentsBayesian Networks for Web Customer Data Mining

4

7

Preprocessing for Text LearningFrom: xxx@sciences.sdsu.eduNewsgroups:comp.graphicsSubject: Need specs on Apple QT

I need to the specs, or at least a very verbose interpretation of the specs, for QuckTime. Technical articles from magazines and references to books would be nice, too.

I also need the specs in a format usable on a Unix or MS-Dos system. I can’t do much with the QuickTime stuff they have on..

specs3unix1

hockey0

computer0

.

space0references1

.

.quicktime2

graphics0

clinton0car0baseball0

8

Text Mining: Data Sets

Usenet Newsgroup Data420 categories 41000 documents for each category 420000 documents in total.

TDT2 Corpus4Target detection and tracking (TDT): NIST4Used 6,169 documents in experiments

5

9

d1 d2 d3 dn

h1 h2 hm

…

…

4 [Chang and Zhang, 2000]4 Input nodes

• Binary values• Represent the existence or

absence of words in documents.

Text Mining:Helmholtz Machine Architecture

−−+

==

∑=

n

jjiji

i

dwbhP

1exp1

1)1(

−−+

==

∑=

m

jjiji

i

hwbdP

1exp1

1)1(

4Latent nodes• Binary values• Extract the underlying causal

structure in the document set. • Capture correlations of the words

in documents.

: recognition weight

: generative weight

10

Text Mining: Learning Helmholtz Machines

4Introduce a recognition network for estimation of a generative network.

4Wake-Sleep Algorithm• Train the recognition and generative models alternately.• Update the weight in network iteratively by simple local delta rule.

( ) ( ) ( )( )

( ) ( )( )∑∑

∑ ∑∑ ∑

=

==

≥

=

=

T

tt

ttt

tt

ttt

T

t t

tt

t QdPQ

QdPQdPD

1)(

)()()(

T

1t )()(

)()()(

1 )(

)()(

)(

|,log

|,log |,log)|log(

α

αα

αθαα

αθααθαθ

))1(( =−=∆

∆+=

jjiij

ijoldij

newij

spssw

www

γ

6

11

Text Mining: Methods

Text Categorization4Train a Helmholtz machine for each category.4Total N machines for N categories.4Once the N machines have been estimated, classification of a test

document proceeds by estimating the likelihood of the document for each machine.

Topic Words Extraction4For the entire document sets, train a Helmholtz machine.4After training, examine the weights of connections from a latent

node to input nodes, that is words.

)]|([logmaxargˆ cdPcCc∈

=

12

4Usenet Newsgroup Data• 20 categories, 1000 documents for each category, 20000 documents in

total.

Text Mining: Categorization Results

7

13

4TDT2 Corpus46,169 documents

imf, monetary, currencies, currency, rupiah, singapore, bailout, traders, markets, thailand, inflation, investors, fund, banks, baht

7

netanyahu, palestinian, arafat, israeli, yasser, kofi, annan, benjamin, palestinians, mideast, gaza,jerusalem, eu, paris, israel

4

India, pakistan, pakistani, delhi, hindu, vajpayee, nuclear, tests, atal, kashmir, indian, janata, bharatiya, islamabad, bihari

5

Suharto, habibie, demonstrators, riots, indonesians, demonstrations, soeharto, resignation, jakarta, rioting, electoral, rallies, wiranto, unrest, megawati

6

pope, cuba, cuban, embargo, castro, lifting, cubans, havana, alan, invasion, reserve, paul, output, vatican, freedom

8

olympics, nagano, olympic, winter, medal, hockey, atheletes, cup, games, slalom, medals, bronze, skating, lillehammer, downhill

3

warplane, airline, saudi, gulf, wright, soldiers, yitzhak, tanks, stealth, sabah, stations, kurds, mordechai, separatist, governor

2

tabacco, smoking, gingrich, newt, trent, republicans, congressional, republicans, attorney, smokers, lawsuit, senate, cigarette, morris, nicotine

1

Text Mining: Topic Words Extraction Results

14

KDD-2000 Web Mining Competition4Data: 465 features over 1700 customers

• Features include friend promotion rate, date visited, weight of items, price of house, discount rate, …

• Data was collected during Jan. 30 – March 30, 2000• Friend promotion was started from Feb. 29 with TV

advertisement. 4Aims: Description of heavy/low spenders

Web Mining: Customer Analysis

8

15

16

9

17

Web Mining: Feature SelectionFeatures selected by various ways [Yang & Zhang, 2000]

V240 (Friend) V229 (Order-Average) V304 (OrderShippingAmtMin.)V368 (Weight Average)V43 (Home Market Value)V377 (NumAcountTemplate Views)+V11 (WhichDoYouWearMostFrequent)V13 (SendEmail)V17 (USState)V45 (VehicleLifeStyle)V68 (RetailActivity)V19 (Date)

V13 (SendEmail)V234 (OrderItemQuantity Sum%HavingDiscountRange(5 . 10))V237 (OrderItemQuantitySum%Having DiscountRange(10.))V240 (Friend)V243 (OrderLineQuantitySum)V245 (OrderLineQuantity Maximum)V304 (OrderShippingAmtMin)V324 (NumLegwearProductViews)V368 (Weight Average)V374 (NumMainTemplateViews)V412 (NumReplenishableStock Views)

V368 (Weight Average)V243 (OrderLine Quantity Sum)V245 (OrderLine Quantity Maximum)F1 = 0.94*V324 + 0.868*V374 + 0.898*V412 F2 = 0.829*V234 + 0.857*V240F3 = -0.795*V237+ 0.778*V304

Discriminant ModelDecision TreeDecisionTree+Factor Analysis

18

Web Mining: Bayesian Nets

Bayesian network4DAG (Directed Acyclic Graph) 4Express dependence relations between variables 4Can use prior knowledge on the data (parameters)

A B C P(A,B,C,D,E) = P(A)P(B|A)P(C|B)

P(D|A,B)P(E|B,C,D)

D E

4Examples of conjugate priors:Dirichlet for multinomial data, Normal-Wishart for normal data

10

19

Web Mining: Results

A Bayesian net for KDD web data

V229 (Order-Average) and V240 (Friend) directly influence V312 (Target)

V19 (Date) was influenced by V240 (Friend) reflecting the TV advertisement.

20

SummaryWe study machine learning methods, such as4Probabilistic neural networks4Evolutionary algorithms4Reinforcement learning

Application areas include4Text mining4Web mining4Bioinformatics (not addressed in this talk)

Recent work focuses on probabilistic graphical models for web/text/bio data mining, including4Bayesian networks4Helmholtz machines4Latent variable models

11

21

22

Bayesian Networks:Architecture

A Bayesian network represents the probabilistic relationships between the variables.

B

G M

L

∏=

=n

iiiXPP

1

)|()( paX pai is the set of parent nodes of Xi.

),|()|()()( ),,|(),|()|()(),,,(

LBMPBGPBPLPGBLMPBLGPLBPLPMGBLP

==

12

23

The network structure represents the naïve Bayes assumption.All nodes are binary.[Hwang & Zhang, 2000]

C

t1 t2 t8754

• C: document class

• ti: ith term

Bayesian Networks:Applications in IR – A Simple BN for Text Classification

24

Dataset4The acq dataset from Reuters-2157848754 terms were selected by TFIDF.4Training data: 8762 documents4Test data: 3009 documents

Parametric Learning4Dirichlet prior assumptions for the network parameter

distributions.

4Parameter distributions are updated with training data.),...,|(Dir)|( 1 iijrijij

hij Sp ααθθ =

),...,|(Dir),|( 11 ii ijrijrijijijh

ij NNSDp ++= ααθθ

Bayesian Networks:Experimental Results

13

25

For training data4Accuracy: 94.28%

For test data4Accuracy: 96.51%

99.3293.76Negative examples75.9896.83Positive examples

Precision (%)Recall (%)

98.6796.88Negative examples89.1795.16Positive examples

Precision (%)Recall (%)

Bayesian Networks:Experimental Results

26

Latent Variable Models:Architecture

Latent Variable Model forTopic Words Extraction and Document Clustering

[Shin & Zhang, 2000]Maximize log-likelihood

4Update , , . 4With EM Algorithm

∑∑ ∑

∑∑

= = =

= =

=

=

N

n

M

m

K

kknkmkmn

N

n

M

mmnmn

zdPzwPzPwdn

wdPwdnL

1 1 1

1 1

)|()|()(log),(

),(log),(

)( kzP )|( km zwP )|( kn zdP

Document Clustering

Topic-Words Extraction

14

27

EM (Expectation-Maximization) Algorithm4Algorithm to maximize pre-defined log-likelihood

Iteration of E-Step and M-Step4E-Step M-Step

∑=

= K

kkmknk

kmknk

mnk

zwPzdPzP

zwPzdPzPwdzP

1)|()|()(

)|()|()(),|(

∑∑

∑

= =

== M

m

N

nmnkmn

N

nmnkmn

km

wdzPwdn

wdzPwdnzwP

1 1

1

),|(),(

),|(),()|(

∑∑

∑

= =

== M

m

N

nmnkmn

M

mmnkmn

kn

wdzPwdn

wdzPwdnzdP

1 1

1

),|(),(

),|(),()|(

∑∑

∑∑

= =

= =

≡

=

M

m

N

nmn

M

m

N

nmnkmnk

wdnR

wdzPwdnR

zP

1 1

1 1

),(

,),|(),(1)(

Latent Variable Models:Learning

28

Topic Words Extraction and Document Clustering with a subset of TREC-8 dataTREC-8 adhoc task data4Documents: DTDS, FR94, FT, FBIS, LATIMES4Topics: 401-450 (401, 434, 439, 450)4401: Foreign Minorities, Germany4434: Estonia, Economy4439: Inventions, Scientific discovery4450: King Hussein, Peace

Latent Variable Models:Applications in IR – Experimental Results

15

29

0.9900.729290003450 (293)0.9270.953920307439 (219)0.6860.996791023820434 (347)0.9300.9022001279401 (300)RecallPrecisionz1z3z4z2Topic (#Docs)

Label (assigned to zk with Maximum P(di|zk) )

jordan, peac, isreal, palestinian, king, isra, arab, meet, talk, husayn, agreem, presid, majesti, negoti, minist, visit, region, arafat, secur, peopl, east, washington, econom, sign, relat, jerusalem, rabin, syria, iraq, …

Cluster 1(z1)

research, technology, develop, mar, materi, system, nuclear, environment, electr, process, product, power, energi, countrol, japan, pollution, structur, chemic, plant, …

Cluster 3(z3)

percent, estonia, bank, state, privat, russian, year, enterprise, trade, million, trade, estonian, econom, countri, govern, compani, foreign, baltic, polish, loan, invest, fund, product, …

Cluster 4(z4)

german, germani, mr, parti, year, foreign, people, countri, govern, asylum, polit, nation, law, minist, europ, state, immigr, democrat, wing, social, turkish, west, east, member, attack, …

Cluster 2(z2 )

Extracted Topic Words (top 35 words with highest P(wj|zk) Topics

Latent Variable Models:Applications in IR – Experimental Results

30

Boosting: Algorithms

A general method of converting rough rules into a highly accurate prediction ruleLearning procedure4 Examine the training set4 Derive a rough rule (weak learner)4 Re-weight the examples in the training set, concentrating on the hard cases for

previous rules4 Repeat T times

Learner Learner Learner Learner

h1 h2 h3 h4),,,( 4321 hhhhf

Importance weightsof training documents

16

31

Boosting: Applied to Text Filtering

Naïve Bayes4Traditional algorithm for text filtering

Boosting naïve Bayes4Using naïve Bayes classifiers as weak learners4 [Kim & Zhang, SIGIR-2000]

)|""(

... )|""()|""()(maxarg

)|()(maxarg

)|()(maxarg

21

1

} ,{

jin

jijijc

n

kjikj

c

jijirrelevantrelevantc

NM

ctroublewP

capproachwPcourwPcP

cwPcP

cdPcPc

j

j

j

=

===

=

=

∏=

∈ Assume independence among terms

32

TREC (Text Retrieval Conference)4 Sponsored by NIST

TREC-7 filtering datasets4Training Documents

• AP articles (1988)• 237 MB, 79919 documents

4Test Documents• AP articles (1989~1990)• 471 MB, 162999 documents

4No. of topics: 50

Example of a document

TREC-8 filtering datasets4Training Documents

• Financial Times (1991~1992)• 167 MB, 64139 documents

4Test Documents• Financial Time (1993~1994)• 382 MB, 140651 documents

4No. of topics: 50

Boosting: Applied to Text Filtering – Experimental Results

17

33

0.509

PIRC

0.500

PIRC

0.505

NTTATTBoostingNTTATTBoosting

0.4600.467

Averaged Scaled F3

0.4520.4610.474

Averaged Scaled F1

0.720

Mer

0.714

PIRC

0.734

PIRCCLBoostingPLT2PLT1Boosting

0.7210.722

Averaged Scaled LF2

0.7130.7120.717

Averaged Scaled LF1

TREC-7

TREC-8

Compared with the state-of-the-art text filtering systems

Boosting: Applied to Text Filtering – Experimental Results

34

Evolutionary Learning:Applications in IR - Web-Document Retrieval

[Kim & Zhang, 2000]

Link Information,HTML Tag Information

chromosomes

<TITLE> <H> <B> … <A>

Retrieval

Fitness

wn…w3w2w1

wn…w3w2w1 wn…w3w2w1 wn…w3w2w1

wn…w3w2w1 wn…w3w2w1

18

35

Evolutionary Learning:Applications in IR – Tag Weighting

Crossover

Truncation selection

Mutation

yn…y3y2y1xn…x3x2x1

chromosome X chromosome Y

chromosome Z (offspring)

zi = (xi + yi ) / 2 w.p. Pc

chromosome X

chromosome X’

change value w.p. Pm

xn…x3x2x1

xn…x3x2x1zn…z3z2z1

36

Evolutionary Learning :Applications in IR - Experimental Results

Datasets4TREC-8 Web Track Data42GB, 247491 web documents (WT2g)4No. of training topics: 10, No. of test topics: 10

Results

19

37

Agent

Environment

Reinforcement Learning:Basic Concept

1. State st Reward rt

3. Reward rt+1

4. State st+1

2. Action at

38

WAIR

User

2. Actioni(modify profile)

Rewardi

1. Statei(user profile)

User profile

...Filtered documents

Document filtering

4. Statei+1

3. Rewardi+1 (relevance feedback)

•retrieve documents •calculate similarity

Reinforcement Learning:Applications in IR - Information Filtering[Seo & Zhang, 2000]

20

39

(%)

Reinforcement Learning:Experimental Results (Explicit Feedback)

40

(%)

Reinforcement Learning:Experimental Results (Implicit Feedback)