1 dynamic pricing of information goods joint work with: gabi koifman, avigdor gal technion onn...
Post on 15-Jan-2016
237 views
TRANSCRIPT
1
Dynamic Pricing of Information Goods
Joint work with:Gabi Koifman, Avigdor GalTechnion
Onn ShehoryIBM Haifa Research Labs
2
Motivation
Rapid growth in electronic commerce The information economy vision (Kephart et al.)
Agents accumulate knowledge, stored in databases
Agents can benefit from trading database tuples
No mechanism for such trade
3
Problem Statement
A mechanism for negotiating database-based information goods requires: – Correctly matching of attributes of database goods– Pricing of (DB-based) information goods
Bob’s Agent Alice’s
Agent
Domain:Stocks Domain:Stocks
ID Quote dateIBM 100 01/01/2002IBM 90 01/02/2002IBM 120 01/03/2002IBM 100 01/04/2002
stockID stockQuote dateIBM 100 01/01/2002IBM 90 01/02/2002
I can sell records to
make profit
I need more information
NOW. Willing to spend 50$ for it.
4
(DB-based) Information goods market vs. traditional market
Negligible marginal cost Uniqueness Pricing Experience goods (Advertising) Delivery Schema/tuple ambiguity
5
Compatibility Evaluation
DB information goods compatibility evaluation can be reduced to the schema mapping problem
A mapping F from S to S’ is a set of |S| pairs (a, a’), a S, a’ S’ {null} and S’=F (S)
μatt(a,a’) is the similarity measure of a, a’
μF is computed based on all μatt in F Utility is based on μF
6
Buyer’s Anxiousness Level
Assumption: willingness to pay is proportional to buyer’s anxiousness
A seller can perform price discrimination across consumers with different anxiousness level
Why should a buyer expose its true anxiousness level?
When discriminating based on TTD (Time To Deliver), learning anxiousness is enabled
(we use Bayesian learning)
7
Market Trends
training period
frame
periodtraining period
frameframe
period
Calc: average supply
Calc: average personal demand
set: reference supply\demand levels
Calc: current supply\demand levels
Re-calc: average supply
Re-calc: average personal demand
Re-set: reference supply\demand levels
8
Utility Evaluation
Distance(seller, buyer) = number of tuples that exist in the seller’s database and not in the buyer’s database
If (distance (seller, buyer)> ) then
proceed with negotiation Computing Distance() is problematic
– Database comparison, or– Zero-knowledge mechanism– Relief: can approximate via statistical measures
9
Pricing Policies
Derivative-Follower (DF) Trial and Error (TA) Personalized Pricing (PP) Market Based Personal Pricing (MBPP)
Posted pricing – DF,TA Price discrimination – PP,MBPP Negotiation based pricing – PP,MBPP
10
Negotiation Participants
DB Exchange agent – Trusted third party – Receives ads, publishes to subscribers
Players: buyers and sellers – Initial database – Buyer: maximize (number of distinct tuples),s.t min(cost)– Seller: maximize (profit)
11
Interaction diagram
Agent 1 Agent 2DBE
Transfer GoodsCloser
Price Negotiation
CounterOffer
CounterOffer
CloseDeal
CloseDeal
TerminateNegotiation
TerminateNegotiation
Seller ProcessOffer Buyer Process
Offer
Market trends learning
AL learning
Utility Evaluation
RequestForDistanc
DistanceReplyCalc Distance(2,1)
Negotiation Model
Contact
RequestToPublishPublishingSeller
WillingToNegotiate
InitialOffer
Compatibility Evaluation
OntoBuilder
μ>T
SafeSignsReplyForQueries
RequestForQueries
Schema-mapping learning μ>T
12
Simulation System
Java language – JMS on J2EE. MS-access database JMS messaging
15 agentsJ2ee serverdatabase
15 agentsJ2ee server
DBE
1:1 service
15 playersJ2ee serverdatabase
15 playersJ2ee serverDBE agent
1:1 service
15 agentsJ2ee serverdatabase
15 agentsJ2ee server
DBE
1:1 service
15 playersJ2ee serverdatabase
15 playersJ2ee serverDBE agent
1:1 service
13
Simulation Participants
Buyers: Anxiousness level Max budget for transaction Distance threshold (0)
Sellers: Current price list Probabilities for anxiousness level
distribution Assumed supply Assumed demand
14
Pricing Policies Evaluation:
System profit /volume Equilibrium
Market settings: Non-competitive market Competitive market
15
System Profit
System Volume
0
50
100
150
200
250
300
Market BasedPricing
PersonalizedPricing
DerivativeFollower
Trial And Error
Non-competitve competitive
Derivativefollower
Trial andError
PersonalizedPricing
MarketBased Pricing
System Profit
01000
2000300040005000
60007000
Market BasedPricing
PersonalizedPricing
DerivativeFollower
Trial AndError
Non-competitve competitive
Derivativefollower
Trial andError
PersonalizedPricing
MarketBased Pricing
16
Equilibrium
deviation policy deviative agent 29 PP1 MBPP 150.22 125.961 TA 23.47 127.731 DF 26.53 112.35
PP agent should deviate to MBPP
MBPP agent should not deviate
deviation policy deviative agent 29 MBPP1 PP 70.02 118.111 TA 20.09 106.521 DF 25.62 116.52
17
Conclusions
We provide mechanism for trading databased-based information goods
Pricing policies that allow negotiation and personalization, perform better than (known in the art) posted pricing
Market based personalized pricing policy performs better than personalized pricing, in terms of stability
18
The End
19
Backup Slides
20
Related Work
Pricing Information Goods– (Varian) price discrimination: an issue when willingness to
pay varies across consumers. Need to: Determine the consumer's willingness to pay Prevent “black market”
Information Economy and Software Agents– (Kephart et al.) The vision– Agent: faster, but less intelligent and flexible
– Effects on Global Economy Multiagent Negotiation
– Protocol, objects, reasoning model (Jennings et al.) Multiagent Learning
– Bayesian learning in negotiation – Zeng and Sycara
21
Future Work
Support buyers that wish to build a database from an initial empty tuples set.
Situations for compatibility that also use auxiliary information.
Suggest techniques that allow a fully automated algorithm.
Additional pricing policies. Suggest a secure algorithm for distance(a,b), with no
use of third trusted party. Allow the buyer to choose a bidding policy that
maximizes its utility under specific market settings.
22
Database-based Information Goods Compatibility Evaluation
Imprecision Mapping Effectiveness Mapping Cost
Evaluation Methodology and Results
23
Compatibility Evaluation (1) :Mapping Imprecision
Evaluation Methodology and Results
Imprecision improvement Using SafeSigns
with insertion control
no change14.2%
no change (zero
imprecision)21.2%
improved50.8%
not improved13.7%
improved
no change
no change (zero imprecision)
not improved
Imprecision improvement using SafeSigns algorithm
improved40.2%
no change (zero
imprecision)21.4%
not improved29.8%
no change8.5%
improved
no change
no change (zero imprecision)
not improved
Improved40.2%
No change8.5%
NotImprove
d29.8%
No Change)0 imprecision(
21.4%No change
14.2%
Not Improved13.7
Improved50.8%
No Change)0 imprecision(
21.2%
Using SafeSigns ability to generate 0-imprecision mappings was doubled!!!
24
Compatibility Evaluation (2) :
Mapping Effectiveness
Mapping EffectivenessEncoutering a specific seller - No Learning
R2 = 0.9824
0.50.520.540.560.580.6
0.620.640.66
1 2 3 4 5 6 7 8 9 10buyer's i encouter with a specific seller
P(s
uff
icie
nt
map
pin
g)
Mapping EffectivenessEncoutering a specific seller
R2 = 0.9806
0.8
0.85
0.9
0.95
1 2 3 4 5 6 7 8 9 10buyer's i encouter with a specific seller
P(s
uff
icie
nt
ma
pp
ing
)P(sufficient mapping)Log. (P(sufficient mapping))
Evaluation Methodology and Results
25
Compatibility Evaluation (3) :
Mapping Cost
Learning Curve no schema-matching learning
010203040506070
1 1001 2001 3001 4001 5001 6001 7001No. Encounter
Nu
m Q
uer
ies
Learning curveAll encouters are considered
0
10
20
30
40
50
1 1001 2001 3001 4001No. encouter
Nu
m Q
uer
ies
number queries
Log. (number queries)
Evaluation Methodology and Results