DCS/CSCI 2350: Social & Economic Networks Sponsored Search Market - - PDF document

dcs csci 2350 social economic networks
SMART_READER_LITE
LIVE PREVIEW

DCS/CSCI 2350: Social & Economic Networks Sponsored Search Market - - PDF document

5/2/18 DCS/CSCI 2350: Social & Economic Networks Sponsored Search Market Reading: Chapter 15 [EK] Video Lecture: http://bit.ly/vcg-gsp Mohammad T . Irfan Final Exam u Saturday, 5/19 at 9am in classroom u Topics included u Epidemics (Ch


slide-1
SLIDE 1

5/2/18 1

DCS/CSCI 2350: Social & Economic Networks

Sponsored Search Market Reading: Chapter 15 [EK] Video Lecture: http://bit.ly/vcg-gsp

Mohammad T . Irfan

Final Exam

u Saturday, 5/19 at 9am in classroom u Topics included

u Epidemics (Ch 21) u Stable marriage problem (Handout—Blackboard) u Matching markets (Ch 10) u Bargaining and power in networks (Ch 12) u WWW structure, link analysis, and web search

(Chapters 13, 14)

u Sponsored search markets (Chapter 15)

u Not included

u Topics not covered in class u Topics not listed above

slide-2
SLIDE 2

5/2/18 2

Questions

u How does Google make money from web

search?

u Connection with auctions and matching markets

Google's yearly revenue

Source: statista.com

slide-3
SLIDE 3

5/2/18 3

Source: statista.com

Google's yearly revenue just from ads

From CHI-2018 Conference Photo: Ishtiaque Ahmed

slide-4
SLIDE 4

5/2/18 4

Agenda

u Brief history u Conventions u Matching markets u Vickrey-Clarke-Groves (VCG) auction u Generalized second-price (GSP) auction

Search market

u Early days

u Yahoo! negotiates deal with advertisers u Shows the ad a certain number of times u Effective?

u Since year ~2000

u Keyword based advertising by Overture (bought

  • ut by Yahoo! in 2003 for $1.63B)

u New market

slide-5
SLIDE 5

5/2/18 5

Conventions

  • 1. Pay per click

u Advertisers pay Google on per click basis u How much?

  • 2. Price setting by auction

u Fixed price for keywords doesn't work. Why? u Main challenge in auction: multiple ad spots

Preliminaries 1

u Search keyword (e.g., pizza) u Multiple ad slots

Slot Ad 1 ? 2 ? 3 ?

slide-6
SLIDE 6

5/2/18 6

Preliminaries 2

u Click-through-rate (CTR) of a slot

u # of clicks/hr for that slot u Assumptions

u Advertisers know CTR u CTR solely depends on the slot, not on the ad

CTR Slot 10 1 5 2 2 3

Preliminaries 3

u Revenue/click

u Advertisers revenue/click u Assumption

u Solely depends on the advertiser, not on the slot u Private information Advertiser Revenue/Click

x 10 y 5 z 2

slide-7
SLIDE 7

5/2/18 7

Matching market

u Will MCP work?

u What will be the valuations of the advertisers? u CTR × revenue/click (Unit: $/hr)

u Doesn't work because Google needs to know

the advertisers' valuations

First-price auction?

u Did not work

u Underbidding u Turbulent market

u Second-price auction for multiple items?

slide-8
SLIDE 8

5/2/18 8

Vickrey-Clarke-Groves (VCG) mechanism

VCG mechanism

u Alternative view of second-price auction

u Outcome maximizes the social welfare u Winner is charged the "harm" he/she causes to the

  • ther bidders

u How to calculate the harm caused by a

bidder?

u Calculate the increase in others' valuations in an

alternative universe where that bidder is not there

slide-9
SLIDE 9

5/2/18 9

VCG price

VCG price: Each bidder pays the harm he/she causes to the other bidders (harm = total amount everyone would've been better off if that bidder were not there)

Steps of VCG mechanism

  • 1. Ask bidders to submit valuations (CTR ×

revenue/click) [valuations are independent and private, but truth-telling is dominant strategy]

  • 2. Choose a socially optimal assignment (e.g., by

MCP auction algorithm ignoring MCP prices)

  • 3. Charge bidders their respective VCG price

u Example

slide-10
SLIDE 10

5/2/18 10

VCG pros and cons

u Pros

u Unique outcome u Maximizes social welfare u Bidders will submit their true valuations

u Cons

u Complex for bidders u Difficult to explain an outcome to bidders

Difference: MCP vs. VCG price

u MCP: posted price for each item (as in

supermarket)

u VCG: personalized price

u VCG price for the same item is different for

different bidder

u Why?

slide-11
SLIDE 11

5/2/18 11

(Superficially) Generalized Second-Price (GSP) Auction

GSP auction

  • 1. Each advertiser bids a single number:

how much they'll pay per click

  • 2. Google gives the i-th slot to the i-th highest

bidder, who pays the bid of (i+1)-st highest bidder

slide-12
SLIDE 12

5/2/18 12

Analysis

u Varian (2007), Edelman, Ostrovsky, Schwartz

(2007)

u Game

u Players: bidders or advertisers u Strategy: bid amount u Payoff: revenue − price paid

u Nash equilibrium

u Nobody has any incentive to change the bid

unilaterally

GSP pros and cons

u Cons (for multiple items or slots)

u Bidders may not be truthful (contrast w/ VCG) u May not optimize social welfare (contrast VCG) u Multiple Nash equilibrium possible (contrast VCG)

u Single item: GSP = VCG = 2nd price auction u Pros

u Nash equilibrium always exists u One Nash eq maximizes social welfare

u Maximizes Google's revenue? Maybe!

slide-13
SLIDE 13

5/2/18 13

Examples

u Bidders gain more by lying u Multiple Nash equilibria u Compare with VCG