CONTENTORIENTED NEGOTIATION IN ECOMMERCE

1
CONTENT-ORIENTED NEGOTIATION IN E-COMMERCE
Bogaziçi University Department of Computer
Engineering

• Reyhan Aydogan
• Thesis Advisor Asst. Prof. Pinar Yolum

2
OUTLINE

• Negotiation Architecture
• Technical Details
• Representation
• Learning Phase
• Similarity Estimation
• Offering Service Mechanism
• Developed System Performance Evaluation
• Discussion

3
Negotiation Architecture
Data Repository (Inventory Information)
?
?
4-Evaluate the offer
1- Request
2-Evaluate Request and Learning
3-Provide Service or Offer alternative
5-Accept or Re-request

ltPreferencesgt ltprice vlow/gt ltspeed
vhigh/gt lt/Preferencesgt
SHARED ONTOLOGY
N-negotiate and provide service
4
Negotiation Challenges

• Representation
• Represent the request and offers
• Learning
• Learn about consumers preferences based on
  requests and counter offers
• Similarity Estimation
• Estimate similarity between the request and
  available services
• Revision
• Revise requests or offers based on incoming
  information

5
Representation

• The request of the consumer and the counter offer
  of the provider are represented as vectors.
• Example domain
• Service Wine
• Service features winery, type of grape, sugar
  level, flavor, body of the wine, color of the
  wine, region
• Example request or offer vector
• (Bancroft, ChardonnayGrape, Dry, Moderate,
  Medium, White, NapaRegion)

winery type of grape sugar level
flavor body color region
6
Learning Phase

• Preferences Relative importance degree of
  features of the service
• Learn preferences over interactions
• Requires incremental learning algorithms
• Learn preferences as concept
• Version Space as an inductive learning technique
• Decision Trees

7
Learning Phase Version Space

• Maintain two extreme hypotheses sets
• The most general hypotheses
• Initially every possible hypotheses is here
• As the consumer rejects offers, this set is
  specialized
• The most specific hypotheses
• Initially empty
• As the consumer makes requests, her requests are
  generalized and kept in this set
• The goal Obtain a single description

8
Modified Version Space

• To support to learn disjunctive concept
• E.g. (red and strong wine) OR (rose and delicate
  wine)
• Extend hypothesis language to support learning
  disjunctive concepts
• Specialize general set minimally
• General set involves all possible hypothesis.
• Generalize specific set minimally
• Specific set only includes positive samples.

9
Decision Trees
FLAVOR
Acceptable Service (Strong and Red)
OR (Moderate and Rose)
Strong Moderate Delicate
COLOR
COLOR
-
Rejectable Service (Strong and Rose)
OR (Moderate and Red) OR (Delicate)
Red Rose Red
Rose

-
-

10
Offering Service

• Random Offering Service
• Offering service considering only the current
  request (SCR)
• Offering Service using Version Space (VS)
• Offering Service using Modified Version Space
  (MVS)
• Offering Service using Decision Trees (DT)

11
Offering Service using MVS

• At the beginning, load all possible services
  (e.g. wine products) to the service list
• After each request, train the MVS with request as
  a positive sample
• If there is an exactly matched service, offer it
• Otherwise,
• Filter the service list with the most general set
  
• Estimate the similarity of each services with the
  most specific set of learning component
• Offer the most similar service

12
Offering Service using DT

• After each request, rebuild the decision tree
• Remove the services from service list, which are
  classified as negative
• Offer the most similar service to the all
  previous and current requests

13
Tverskys Similarity Measure

• Terms
• Common number of matched attributes
• Different number of unmatched attributes
• a and ß WeightsHere a is equal to ß
• Example
• S1 ( Full, Strong, Red )
• S2 (Full, Delicate, Rose) SMs1s2 1 / 3

14
Architectural Setup

• Implementation in Java
• Ontology language OWL
• Ontology ReasonerJena2
• Ontology
• Shared ontology modified version Wine ontology
• Producers service ontology WineStock
  extension of wine ontology

15
Evaluating The Learning Phase

• Criteria Number of iterations for consensus
• Five systems are compared
• Similarity with Modified Version Space (SMVS)
• System using Decision Trees (DT)
• Similarity with Version Space (SVS)
• Similarity with Current Request (SCR)
• Random Offering (Random)
• Use five scenarios
• Run five times and take average of runs
• Inventory that contains 19 available services

16
Evaluating The Learning Phase Cont.

• Scenario 1
• Preference of consumer Any wine whose sugar
  level is dry
• Availability in producers inventory 15 products
  
• Scenario 2
• Preference of consumer Any wine, which is red
  and dry
• Availability in producers inventory Eight
  products
• Scenario 3
• Preference of consumer Any wine, which is red
  ,dry and moderate
• Availability in producers inventory Four
  products
• Scenario 4
• Preference of consumer Any wine, which is strong
  and red
• Availability in producers inventory Two
  products
• Scenario 5
• Preference of consumer Any wine whose flavor is
  strong and color is red or rose
• Availability in producers inventory Three
  products

17
Evaluating The Learning Phase Cont.

• Average number of iterations for five scenarios

18
Similarity Measure

• Tverskys Similarity Measure
• Proposed Semantic Similarity Measure (RP)
• Resniks Semantic Similarity Measure
• Lins Semantic Similarity Measure
• Wu Palmers Semantic Similarity Measure

19
RP Semantic Similarity

• Parent versus Grandparent
• Reddish Color is more similar than WineColor to
  Rose
• Parent versus Sibling
• WineColor is more similar than ReddishColor to
  White
• Sibling versus Grandparent
• Red is more similar than WineColor to Rose

20
RP Semantic Similarity Cont.

• Start the similarity with one at the node
  containing the first concept and decrease it by
  some constant at each level
• Assume
• m is the constant for parents
• n is the constant for siblings

21
RP Semantic Similarity Sample

• Rose-ReddishColor
• 1 (2/3) 0.67
• Rose-Red
• 1 (4/7) 0.57
• Rose-WineColor
• 1 (2/3)(2/3) 0.45
• Rose-Thing
• 1(2/3)(2/3)(2/3) 0.30
• Rose-White
• 1(4/7)(2/3) 0.38

• Assume
• m2/3 and n4/7

22
Evaluating Similarity Metrics

• Scenario 1-7 use dataset1 (19 services)
• Scenario 8-10 use dataset2 (50 services)
• Scenario 6-10 consider the hierarchical relation
  in preferences
• Sample scenario 9
• expensive red wine, which is located around
  California region or cheap white wine, which is
  located in around Texas region.

23
Evaluating Similarity Metric Cont.

• Average number of iterations for ten scenarios

24
General Results

• Learning preferences shorten the negotiation
  duration
• Usage of semantic similarity increases the
  performance when preferences are concerned
• Using Modified Version Space or Decision Trees
  results in reasonable results.

25
Contributions of thesis

• A multi-issue negotiation mechanism based on the
  content of the service
• Usage of ontologies so work with semantics
• Extension of CEA Algorithm for disjunctive
  concepts
• A new semantic similarity measure

26
Future Work

• Modeling producers preferences and business
  policy
• The producer may prefer to provide some services
  over others
• Integration of learning with ontology reasoning