04/01/20031 Project and Product Selection by He Jiang Department of Management University of Utah April 1 st, 2003.

04/01/2003 1

Project and Product Selection

by He Jiang

Department of ManagementUniversity of Utah

April 1st, 2003

04/01/2003 2

Outline

• On Integrating Catalogs

• A Hierarchical Constraint Satisfaction Approach to Product Selection for Electronic Shopping Support

• A Multiple Attribute Utility Theory Approach to Ranking and Selection

04/01/2003 3

On Integrating Catalogs

Rakesh Agrawal and Ramakrishnan Srikant

IBM Almaden Research Center

04/01/2003 4

Summary

• Problem: integrating documents from different sources into a master catalog.

• Gaps: Many data sources have their own categorizations; implicit similarity information in these source catalogs may be ignored.

• Approaches: Naïve Bayes classification• Contribution: classification accuracy can be

improved by incorporate the implicit similarity information present in these source categorizations

04/01/2003 5

Problem—Why Integration?

• B2C shops need to integrate catalogs from multiple vendors ( Amazon);

• B2B portals merged into one company (Chipcenter & Questlink eChips);

• Information portals categorize documents into categories (Google & Yahoo!).

• Corporate portals Merge intra-company and external information into a uniform categorization

04/01/2003 6

Problem Identification—Model Building

• Problem identification: classification problem.

• Master catalog M with categories C1, C2, …, Cn;

• Source catalog N with categories S1, S2, …, Sm;

• Merge documents in N into M.

04/01/2003 7

Question

How to Integrate?

04/01/2003 8

Straightforward Approach:• Completely ignore N’s categorization, put each of N’s

product into M’s category according to M’s classification rule.

04/01/2003 9

Enhanced Approach

• incorporate the implicit categorization information present in N into M.

04/01/2003 10

Assumptions and Limitations

• M and N may are homogeneous and have significant overlap;

• M and N use the same vocabularies (Larkey, 1999).

• Catalog hierarchies is flattened and is treated as a set of categories(Good 1965 & Chakrabarti 1997)

• Different hierarchy levels (if M>N, can help distinguish categories that M doesn’t have; if N>M, NBHC can be applied.

04/01/2003 11

Related Works and Gaps

• Naïve-Bayes classifiers are accurate and fast(Chakrabarti et al 1997, …), so we choose Bayesian model;

• Folder systems such as email routing(Agrawal et al, 2000,…), action predicting(Maes, 1994 & Payne et al, 1997), query organizing using text clustering(Sahami et al, 1998) and filings transferring(Dolin et al 1999); But none of this systems address the task of merging hierarchies

• The Athena system includes the facility of reorganizing folder hierarchy into a new hierarchy (Agrawal et al, 2000); But no information from the old hierarchy is used in either building the model or routing the documents.

04/01/2003 12

Straightforward Approach

othereach oft independen are din words:Assumption

)|Pr()|Pr(•

dataset in the documents ofnumber Total

CCategory in documents ofNumber )Pr(•

)Pr(

)|Pr()Pr()|Pr(

i

dt

ii

i

iii

CtCd

C

d

CdCdC

:ddocument agiven

Ccategory ofy probabilitPosterior :classifier Bayes Naïve i

04/01/2003 13

t ii

i

i

ii

tCnCn

tCn

VCn

tCnCt

),()(

y. vocabular theof size theis |V|

Cicategory in documents

in t wordof soccurrence ofnumber the—),(

0 ,||)(

),()|Pr(

Straightforward Approach—Continued

04/01/2003 14

Enhanced Bayes Classification

M.in Ccategory in documents ofnumber theis |C| where

) )Cin be topredicted Sin docs of(Number |C(|

)Cin be topredicated Sin documents ofNumber (|C|

Sin documents ofnumber Total

Cin be topredicated Sin documents ofNumber )|Pr(

Nin category a denotes S where

)|Pr(

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

ii

j

n

1j j

ii

i

SC

Sd

CdSCSdC

i

iii

04/01/2003 15

Effect of Weight on Accuracy

• Weight can make difference for a given M and N; Tune set method to select a good value for the weight.

•

•

in which the document will be correctly classified or will never be correctly classified

• The highest possible accuracy achievable with the enhanced algorithm is no worse than what can be achieved with the basic algorithm.

ff then , if 0, , weight ofpair given any For :1 Theorem21 xx2121

), ,( interval aexist thered,document each For :2 Theorem 21

04/01/2003 16

Experimental Results—Data Sets Used

• Synthetic catalog: deriving source catalog N from M using different distributions(e.g. Gaussian).

• Real Catalog: two real-world catalogs that have some common documents; treat the first catalog minus the common documents as M, the remaining documents in the second catalog as N;

04/01/2003 17

Experimental Results

04/01/2003 18

Experimental Results

04/01/2003 19

Experimental Results

04/01/2003 20

Experimental Results—Catalog Size

04/01/2003 21

Experimental Results—Catalog Size

04/01/2003 22

Contributions and Future Research Directions

• Contributions: enhancing the standard Naive Bayes classification by incorporating the category information of the source catalogs; the highest accuracy of the enhanced technique can be no worse than that can be achieved by standard Naïve Bayes classification.

• Future research: using other classifiers such as SVM to incorporating the implicit information of N requires further work

04/01/2003 23

A Hierarchical Constraint Satisfaction Approach to Product Selection for

Electronic Shopping Support

Young U. Ryu

IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and humans

Vol. 29, No. 6, November 1999

04/01/2003 24

Summary

• Problem: proposing a product selection mechanism for electronic shopping support;

• Approach: hierarchical constraint satisfaction (HCS) approach

• Gap: simple taxonomy hierarchy(STH) approach is flawed in that the the search is conducted on a single generic product hierarchy;

• HCS is more powerful and flexible than STH.

04/01/2003 25

Simple taxonomy Hierarchy Approach

04/01/2003 26

Question

• 1. How do we search for a sugar-free decaffeinated cola?

• 2. If there isn’t a cola that satisfy all the requirements, i.e., cola, sugar-free and decaffeinated. what’s your recommendation?

04/01/2003 27

Gaps

• Search is conducted on a single generic product hierarchy;

• There may exist a product that cannot satisfy all the constraints;

• A product may be evaluated to be better than another while there is no big differences between these two products.

04/01/2003 28

Hierarchical Constraint Satisfaction Approach

• Constraint Satisfaction: a methodology determining assignments of values to variables that are consistent with given constraint;

• Hierarchical Constraint Satisfaction: an extension of STH which minimizes the the satisfaction errors of hierarchically organized constraints based on their importance;

• Value of HCS: can be applied to cases in which there isn’t a solution that is consistent with given constraints due to conflicting constraints.

04/01/2003 29

Concepts Introduced

• Constraint domain transformation: transformation of a Boolean constraint to a arithmetic constraint;

• Tree domain: is one whose elements are structured as a tree; thus can be handled more flexibly;

• Indifference interval: overcome a shortcoming of hierarchical reasoning when the difference between two alternatives is small;

04/01/2003 30

Constraint Satisfaction Error

• Measures the degree of satisfaction of an arithmetic constrain c by the constraint satisfaction error function

• for Boolean constraint, transform them into arithmetic constraints;

• e.g.

),( ic

......20045

1805510545

,

20045

18055

320

,

),,(

3

2

1

3

2

1

321

x

xx

ifgreen

x

x

x

ifred

xxxf

04/01/2003 31

Hierarchical reasoning and indifference interval

ijevery for ),'(c)(c

and ),'(c)(c

n,i somefor if

'product another n better tha is product A

j, j,

j, i,

xx

ijevery for ,|)'(c)(c|

and ,)(c)'(c

:ionconsiderat into interval ceindifferen take

j, j,

i, i,

j

i

i

c

c

c

04/01/2003 32

Constraint Hierarchies

iijijiji

Ccijiji

h

h

CcxcwxC

xcwxC

cc

CH

iij

:),(max),(or

),(),(

iserror on satisfacti sconstraint aggregatedthen

c'.than important more is c constraint that means '

Con relation orderingk strict wea a is

sconstraint ofset a is C where

,pair a ishierarchy constraintA

h

04/01/2003 33

Example

• Shopping for wipes products using hierarchical constraint satisfaction approach. Each product is described by the following attributes:

• Cost: cents per sheet

• Add-on materials: “baking soda”, “aloe vera”, …;

• Strength: measured by pressure(psi) that breaks a sheet;

• Dispenser type: “box”, “pop-up”;

• Added artificial scent: unscented, natural aloe scented, natural jasmine scented and chemical perfume scented;

• Product purpose: “general purpose”, “diaper change”.

04/01/2003 34

Example—Result

04/01/2003 35

Contributions and Future Research Directions

• Contribution: the product search mechanism is viewed as a satisfaction problem of hierarchically organized constraints over product attributes, thus it is more powerful and flexible than product selection based on a single product taxonomy hierarchy.

• Future research: Purchasing requirement specification or constraint hierarchy elicitation; complete prototype implementation of the HCS approach; actual purchasing/sales transaction based on speech –act theory, illocutionary logic and inter-organizational activity coordination.

04/01/2003 36

A Multiple Attribute Utility Theory Approach to ranking and Selection

John Butler, Douglas J. Morrice and

Peter W. Mullarkey

Management Science, Vol. 47, No. 6, June 2001

04/01/2003 37

Summary• Problem: developing a ranking and selection

procedure for making comparison of systems that have multiple performance measures;

• Approach: combining Multiple Attribute Utility Theory (MAUT) and statistical ranking and selection (R&S) using indifference zone;

• Gaps: costing approach is flawed in that accurate cost data may not be available, and it may be difficult to measure performance using costs..

• Advantages: rigorous; close to business practice; simpler to implement; can estimate the number of simulations required; can assess the relative importance of criteria

04/01/2003 38

Gaps

• Most of the R&S literature focused on procedures that reduce the multivariate performance measures to a scalar performances measure problem, but these procedures may have some disadvantages, e.g. accurate cost data may not be available; it maybe difficult to accurately attach a dollar value to intangible variables;

• Current techniques may require a complicated step of estimating a covariance matrix(Gupta & Panchapakesan 1979);

• Previous work doesn’t provide an approach to estimate the number of simulations required to select the best configurations with a high level of probability(Andijani 1998, Kim & Lin 1999).

• Previous work lacks a trade-off mechanism that allows the decision maker to combine disparate performance measures.

04/01/2003 39

Assumptions

• Decision maker’s preferences are accurately represented ( Clemen 1991, Keeney & Raiffa 1976);

• Performance measures that is converted to “utils” can be converted to meaningful unit by choosing an invertible utility function;

• There is a indifference zone for the decision maker on all the performance measures;

04/01/2003 40

General Outline of the Procedure

Construct MAU Model

Run Simulator

Assess Indifference Zone

Simulation Output Vector

Apply MAU Model and Scalar Based R&S Procedures

Sensitivity Analysis on MAU Weights

Assess Utility Functions

Assess Weights

Utility Exchange

04/01/2003 41

s.preferenceon m and j i, attributes

between n interactio therepresent that constants scaling the

are wi, measurefor weight theis,10 i, measureover

functionutility sigle a is ,1(.)0 measures, eperformanc

over variablesrandom of vestor a is )...,( Where

)()...()(...

)()()(

)()()()(

ijm

21

2211...123

1

11

i

i

n

nnn

n

i ij ijmmmjjiiijm

jj

n

i ijiiij

n

iiii

w

u

XXXX

XuXuXuw

XuXuXuw

XuXuwXuwXu

Multilinear Utility Function

04/01/2003 42

Multiplicative MAU Model

.w1- and ,10 Where

)()...()(...

)()()(

)()()()(

)](1[)(1

then t,independenutility mutual If

2211...123

1

11

1

i

nnn

n

i ij ijmmmjjiiijm

jj

n

i ijiiij

n

iiii

i

n

i ii

w

XuXuXuw

XuXuXuw

XuXuwXuwXu

XuwwXwu

04/01/2003 43

Additive MAU Model

• If mutual utility additive independent, then

• Example for additive independence:

n

iiii XuwXu

1

)()(

0.5probility with ),(x

0.5probility with ),(x B

0.5probility with ),(x

0.5probility with ),(x

*

*

**

21

*

2*

1

2*

1*

21

x

x

x

xA

04/01/2003 44

Single Attribute Utility Function Used

•

• Methods for assigning weights: trade-off method; analytical hierarchy process (AHP).

i. measurefor constants scaling are and and

ancerisk toler smaker'decision theis where

)( )(

ii

i

RTxiiii

BA

RT

eBAxu ii

04/01/2003 45

Question

• What’s the benefit of using this function?

04/01/2003 46

R&S Experimental Set-up

• Correct Selection (CS): the R&S procedure accurately identifies the configuration with largest expected utility .

• Two stage indifference zone procedure for R&S.

)]([ ][KXuE

10 and ,11 where

)]E[u(X-

)]E[u(X whenever }{

*

*1]-[K

[K]*

*P/K)(

PCSP

04/01/2003 47

Selection of

• A Utility Exchange ApproachTable 1 Alternatives by Measures Matrix for Car Selection

Table 2 Equivalent Hypothetical Cars

*

Alternative Cost Horsepower Harmony 17,000 160 Starburst 17,000 125 Keyo 15,200 100 Palomino 18,500 160

Alternative Cost Horsepower Harmony 17,000 140 Starburst 17,000 140 Keyo 15,200 140 Palomino 18,500 160

04/01/2003 48

Question Again

• Does it mean that the 20 horsepower is worth $1,200?

04/01/2003 49

Selection of

*

ikiik

kk

k

cxux

Q

Qxuux

x

)( such that specified is ' where

))(

('

is ' of level computingfor expressionan

function,utility r Multilinea for the :1n Propositio •

'1

2

1111

1

function. AMUfor K, , 2, ,1k

k,ion configurat and 1 measure standard for the

))(var())(var(

:hold iprelationsh following The :2n Propositio •

22

'11

Q

XuXu K

K

04/01/2003 50

Establishing the Indifference Zone

• Curve dividing the indifference and preference zone:

1

1]1[

)(ln1

*1

11]1[1][RT

CE

KK

K

eB

RTCECE

04/01/2003 51

Different for ZoneceIndifferen theofzation Characteri *

04/01/2003 52

• Example:

Different for ZoneceIndifferen theofzation Characteri *

04/01/2003 53

Application of the Procedure—Case Description

• Case example: Land Seismic Survey;• Performance measures: survey cost; survey

duration; utilization of the four crews; • Relationship of the crews:

04/01/2003 54

Application of the Procedure—Results

04/01/2003 55

Application of the Procedure—Results

04/01/2003 56

Application of the Procedure—Sensitivity Analysis to Weight

04/01/2003 57

Contributions and Future Research Directions

• Contribution: provides a formal procedure that can be applied to realistic problems; presents a scalar performance measure that can summarize performance on multiple criteria, including nonlinear preference functions and the relative importance of the measures;

• Future research: combine MAU theory with the work of Chen et al; extend the MAU methodology with Chick and Inoue’s work to include their Bayesian technique and relieve some of the computational burden of all R&S procedure; combine the work in this paper with R&S procedures designed facilitate variance reduction through the use of common random numbers (See Matejcik and Nelson 1995 and Goldman and Nelson 1998).