A glimpse to sponsored search auctions Maria Serna Fall 2016 - - PowerPoint PPT Presentation

a glimpse to sponsored search auctions
SMART_READER_LITE
LIVE PREVIEW

A glimpse to sponsored search auctions Maria Serna Fall 2016 - - PowerPoint PPT Presentation

Keyword Auctions Basic model GSP auction A glimpse to sponsored search auctions Maria Serna Fall 2016 AGT-MIRI Sponsored search Keyword Auctions Basic model GSP auction 1 Keyword Auctions 2 Basic model 3 GSP auction AGT-MIRI Sponsored


slide-1
SLIDE 1

Keyword Auctions Basic model GSP auction

A glimpse to sponsored search auctions

Maria Serna Fall 2016

AGT-MIRI Sponsored search

slide-2
SLIDE 2

Keyword Auctions Basic model GSP auction

1 Keyword Auctions 2 Basic model 3 GSP auction

AGT-MIRI Sponsored search

slide-3
SLIDE 3

Keyword Auctions Basic model GSP auction

Keyword Auctions/ Sponsored search

AGT-MIRI Sponsored search

slide-4
SLIDE 4

Keyword Auctions Basic model GSP auction

Keyword Auctions/ Sponsored search

Advertiser submit bids for keywords

AGT-MIRI Sponsored search

slide-5
SLIDE 5

Keyword Auctions Basic model GSP auction

Keyword Auctions/ Sponsored search

Advertiser submit bids for keywords

Offer a dollar payment per click. Alternatives: price per impression, or per conversion.

AGT-MIRI Sponsored search

slide-6
SLIDE 6

Keyword Auctions Basic model GSP auction

Keyword Auctions/ Sponsored search

Advertiser submit bids for keywords

Offer a dollar payment per click. Alternatives: price per impression, or per conversion.

Separate auction for every query

AGT-MIRI Sponsored search

slide-7
SLIDE 7

Keyword Auctions Basic model GSP auction

Keyword Auctions/ Sponsored search

Advertiser submit bids for keywords

Offer a dollar payment per click. Alternatives: price per impression, or per conversion.

Separate auction for every query

Positions awarded by some mechanism. Advertisers get a price per click.

AGT-MIRI Sponsored search

slide-8
SLIDE 8

Keyword Auctions Basic model GSP auction

Keyword Auctions/ Sponsored search

Advertiser submit bids for keywords

Offer a dollar payment per click. Alternatives: price per impression, or per conversion.

Separate auction for every query

Positions awarded by some mechanism. Advertisers get a price per click.

Some new features

AGT-MIRI Sponsored search

slide-9
SLIDE 9

Keyword Auctions Basic model GSP auction

Keyword Auctions/ Sponsored search

Advertiser submit bids for keywords

Offer a dollar payment per click. Alternatives: price per impression, or per conversion.

Separate auction for every query

Positions awarded by some mechanism. Advertisers get a price per click.

Some new features

Multiple positions, but advertisers submit only a single bid. Search is highly targeted, and transaction oriented.

AGT-MIRI Sponsored search

slide-10
SLIDE 10

Keyword Auctions Basic model GSP auction

Keyword auctions: evolution

AGT-MIRI Sponsored search

slide-11
SLIDE 11

Keyword Auctions Basic model GSP auction

Keyword auctions: evolution

Pre-1994: advertising sold on a per-impression basis, traditional direct sales to advertisers.

AGT-MIRI Sponsored search

slide-12
SLIDE 12

Keyword Auctions Basic model GSP auction

Keyword auctions: evolution

Pre-1994: advertising sold on a per-impression basis, traditional direct sales to advertisers. 1994: Overture (then GoTo) allows advertisers to bid for keywords, offering some amount per click. Advertisers pay their bids.

AGT-MIRI Sponsored search

slide-13
SLIDE 13

Keyword Auctions Basic model GSP auction

Keyword auctions: evolution

Pre-1994: advertising sold on a per-impression basis, traditional direct sales to advertisers. 1994: Overture (then GoTo) allows advertisers to bid for keywords, offering some amount per click. Advertisers pay their bids. Late 1990s: Yahoo! and MSN adopt Overture, but mechanism proves unstable. Advertisers constantly change bids to avoid paying more than necessary.

AGT-MIRI Sponsored search

slide-14
SLIDE 14

Keyword Auctions Basic model GSP auction

Keyword auctions: evolution

Pre-1994: advertising sold on a per-impression basis, traditional direct sales to advertisers. 1994: Overture (then GoTo) allows advertisers to bid for keywords, offering some amount per click. Advertisers pay their bids. Late 1990s: Yahoo! and MSN adopt Overture, but mechanism proves unstable. Advertisers constantly change bids to avoid paying more than necessary. 2002: Google modifies keyword auction to have advertisers pay minimum amount necessary to maintain their position (GSP) - followed by Yahoo! and MSN.

AGT-MIRI Sponsored search

slide-15
SLIDE 15

Keyword Auctions Basic model GSP auction

1 Keyword Auctions 2 Basic model 3 GSP auction

AGT-MIRI Sponsored search

slide-16
SLIDE 16

Keyword Auctions Basic model GSP auction

A basic model: the setting

AGT-MIRI Sponsored search

slide-17
SLIDE 17

Keyword Auctions Basic model GSP auction

A basic model: the setting

We consider an auction with n advertisers and n slots.

AGT-MIRI Sponsored search

slide-18
SLIDE 18

Keyword Auctions Basic model GSP auction

A basic model: the setting

We consider an auction with n advertisers and n slots. Slots have associated fixed and public click-through-rates α1 ≥ α2 ≥ · · · ≥ αn.

AGT-MIRI Sponsored search

slide-19
SLIDE 19

Keyword Auctions Basic model GSP auction

A basic model: the setting

We consider an auction with n advertisers and n slots. Slots have associated fixed and public click-through-rates α1 ≥ α2 ≥ · · · ≥ αn. By setting some of them = 0, the case with k < n slots is included.

AGT-MIRI Sponsored search

slide-20
SLIDE 20

Keyword Auctions Basic model GSP auction

A basic model: the setting

We consider an auction with n advertisers and n slots. Slots have associated fixed and public click-through-rates α1 ≥ α2 ≥ · · · ≥ αn. By setting some of them = 0, the case with k < n slots is included. Advertiser i submits a bid bi, the amount he is willing to pay for a click.

AGT-MIRI Sponsored search

slide-21
SLIDE 21

Keyword Auctions Basic model GSP auction

A basic model: outcome

AGT-MIRI Sponsored search

slide-22
SLIDE 22

Keyword Auctions Basic model GSP auction

A basic model: outcome

An outcome is an assignment of advertisers to slots and of payments per click.

AGT-MIRI Sponsored search

slide-23
SLIDE 23

Keyword Auctions Basic model GSP auction

A basic model: outcome

An outcome is an assignment of advertisers to slots and of payments per click. An assignment can be model as a permutation π. π(j) is the advertiser assigned to slot j.

AGT-MIRI Sponsored search

slide-24
SLIDE 24

Keyword Auctions Basic model GSP auction

A basic model: outcome

An outcome is an assignment of advertisers to slots and of payments per click. An assignment can be model as a permutation π. π(j) is the advertiser assigned to slot j. A payment vector p, where pi is the price per click for advertiser i.

AGT-MIRI Sponsored search

slide-25
SLIDE 25

Keyword Auctions Basic model GSP auction

A basic model: outcome

An outcome is an assignment of advertisers to slots and of payments per click. An assignment can be model as a permutation π. π(j) is the advertiser assigned to slot j. A payment vector p, where pi is the price per click for advertiser i. The benefit per click is assumed to be independent of the slot.

AGT-MIRI Sponsored search

slide-26
SLIDE 26

Keyword Auctions Basic model GSP auction

A basic model: parameters

AGT-MIRI Sponsored search

slide-27
SLIDE 27

Keyword Auctions Basic model GSP auction

A basic model: parameters

Each advertiser i has a private value vi, his value per click. The sequence v = (v1, ..., vn) is the valuation profile.

AGT-MIRI Sponsored search

slide-28
SLIDE 28

Keyword Auctions Basic model GSP auction

A basic model: parameters

Each advertiser i has a private value vi, his value per click. The sequence v = (v1, ..., vn) is the valuation profile. Each advertiser i has a quality factor γi that reflects the clickability of its ad.

AGT-MIRI Sponsored search

slide-29
SLIDE 29

Keyword Auctions Basic model GSP auction

A basic model: parameters

Each advertiser i has a private value vi, his value per click. The sequence v = (v1, ..., vn) is the valuation profile. Each advertiser i has a quality factor γi that reflects the clickability of its ad. In the simplest model γi = 1.

AGT-MIRI Sponsored search

slide-30
SLIDE 30

Keyword Auctions Basic model GSP auction

A basic model: utilities

AGT-MIRI Sponsored search

slide-31
SLIDE 31

Keyword Auctions Basic model GSP auction

A basic model: utilities

When advertiser i is assigned to the j-th slot, she gets αjγi clicks.

AGT-MIRI Sponsored search

slide-32
SLIDE 32

Keyword Auctions Basic model GSP auction

A basic model: utilities

When advertiser i is assigned to the j-th slot, she gets αjγi clicks. If advertiser i is assigned to slot j at a price of pi per click then her utility is ui = αjγi(vi − pi), which is the number of clicks received times profit per click.

AGT-MIRI Sponsored search

slide-33
SLIDE 33

Keyword Auctions Basic model GSP auction

A basic model: welfare

AGT-MIRI Sponsored search

slide-34
SLIDE 34

Keyword Auctions Basic model GSP auction

A basic model: welfare

The social welfare of an outcome π is the total value of the solution for the participants, including the auctioneer.

AGT-MIRI Sponsored search

slide-35
SLIDE 35

Keyword Auctions Basic model GSP auction

A basic model: welfare

The social welfare of an outcome π is the total value of the solution for the participants, including the auctioneer. SW (p, π, v, γ) =

n

  • i=1

απ−1(i)γi(vi − pi) +

n

  • j=1

αjγπ(j)pπ(j) =

n

  • j=1

αjγπ(j)vπ(j)

AGT-MIRI Sponsored search

slide-36
SLIDE 36

Keyword Auctions Basic model GSP auction

A basic model: welfare

The social welfare of an outcome π is the total value of the solution for the participants, including the auctioneer. SW (p, π, v, γ) =

n

  • i=1

απ−1(i)γi(vi − pi) +

n

  • j=1

αjγπ(j)pπ(j) =

n

  • j=1

αjγπ(j)vπ(j) The social welfare is independent of the payments and the bids! SW (π, v, γ)

AGT-MIRI Sponsored search

slide-37
SLIDE 37

Keyword Auctions Basic model GSP auction

VCR mechanism

AGT-MIRI Sponsored search

slide-38
SLIDE 38

Keyword Auctions Basic model GSP auction

VCR mechanism

The optimal social welfare is Opt(v, γ) = max

π

SW (π, v, γ).

AGT-MIRI Sponsored search

slide-39
SLIDE 39

Keyword Auctions Basic model GSP auction

VCR mechanism

The optimal social welfare is Opt(v, γ) = max

π

SW (π, v, γ). The efficient outcome sorts advertisers by their effective values γivi, and assigns them to slots in this order.

AGT-MIRI Sponsored search

slide-40
SLIDE 40

Keyword Auctions Basic model GSP auction

VCR mechanism

The optimal social welfare is Opt(v, γ) = max

π

SW (π, v, γ). The efficient outcome sorts advertisers by their effective values γivi, and assigns them to slots in this order. We can design the associated VGC mechanism in which truthfulness is a dominant strategy.

AGT-MIRI Sponsored search

slide-41
SLIDE 41

Keyword Auctions Basic model GSP auction

VCR mechanism

The optimal social welfare is Opt(v, γ) = max

π

SW (π, v, γ). The efficient outcome sorts advertisers by their effective values γivi, and assigns them to slots in this order. We can design the associated VGC mechanism in which truthfulness is a dominant strategy. Exercise: what would be the prices in the VCG auction?

AGT-MIRI Sponsored search

slide-42
SLIDE 42

Keyword Auctions Basic model GSP auction

1 Keyword Auctions 2 Basic model 3 GSP auction

AGT-MIRI Sponsored search

slide-43
SLIDE 43

Keyword Auctions Basic model GSP auction

GSP: Generalized Second Price Auction

AGT-MIRI Sponsored search

slide-44
SLIDE 44

Keyword Auctions Basic model GSP auction

GSP: Generalized Second Price Auction

Players are asked to submit a bid, which is his reported valuation.

AGT-MIRI Sponsored search

slide-45
SLIDE 45

Keyword Auctions Basic model GSP auction

GSP: Generalized Second Price Auction

Players are asked to submit a bid, which is his reported valuation. Given a bid profile b, we define the effective bid of advertiser i to be γibi

AGT-MIRI Sponsored search

slide-46
SLIDE 46

Keyword Auctions Basic model GSP auction

GSP: Generalized Second Price Auction

Players are asked to submit a bid, which is his reported valuation. Given a bid profile b, we define the effective bid of advertiser i to be γibi which is her bid modified by her quality factor, analogous to the effective value defined above.

AGT-MIRI Sponsored search

slide-47
SLIDE 47

Keyword Auctions Basic model GSP auction

GSP: Generalized Second Price Auction

Players are asked to submit a bid, which is his reported valuation. Given a bid profile b, we define the effective bid of advertiser i to be γibi which is her bid modified by her quality factor, analogous to the effective value defined above. The auctioneer sets π(k) to be the advertiser with the kth highest effective bid (breaking ties arbitrarily).

AGT-MIRI Sponsored search

slide-48
SLIDE 48

Keyword Auctions Basic model GSP auction

GSP: Generalized Second Price Auction

Players are asked to submit a bid, which is his reported valuation. Given a bid profile b, we define the effective bid of advertiser i to be γibi which is her bid modified by her quality factor, analogous to the effective value defined above. The auctioneer sets π(k) to be the advertiser with the kth highest effective bid (breaking ties arbitrarily). That is, the GSP mechanism assigns slots with higher click-through-rate to advertisers with higher effective bids.

AGT-MIRI Sponsored search

slide-49
SLIDE 49

Keyword Auctions Basic model GSP auction

GSP:pricing

AGT-MIRI Sponsored search

slide-50
SLIDE 50

Keyword Auctions Basic model GSP auction

GSP:pricing

Prices per click are set as the smallest bid that guarantees the advertiser the same slot.

AGT-MIRI Sponsored search

slide-51
SLIDE 51

Keyword Auctions Basic model GSP auction

GSP:pricing

Prices per click are set as the smallest bid that guarantees the advertiser the same slot. When advertiser i is assigned to slot k (that is, when π(k) = i), this critical value is defined as pi = γπ(k+1) γi bπ(k+1). where we take bn+1 = 0.

AGT-MIRI Sponsored search

slide-52
SLIDE 52

Keyword Auctions Basic model GSP auction

GSP:pricing

Prices per click are set as the smallest bid that guarantees the advertiser the same slot. When advertiser i is assigned to slot k (that is, when π(k) = i), this critical value is defined as pi = γπ(k+1) γi bπ(k+1). where we take bn+1 = 0. In the case γi = 1, for each i, pi = bπ(k+1).

AGT-MIRI Sponsored search

slide-53
SLIDE 53

Keyword Auctions Basic model GSP auction

GSP:utility

AGT-MIRI Sponsored search

slide-54
SLIDE 54

Keyword Auctions Basic model GSP auction

GSP:utility

ui(b, γ) is the utility derived by advertiser i from the GSP mechanism when advertisers bid according to b: ui(b, γ) =απ−1(i)γi(vi − pi) = απ−1(i)[γivi − γπ(π−1(i)+1)bπ(π−1(i)+1)].

AGT-MIRI Sponsored search

slide-55
SLIDE 55

Keyword Auctions Basic model GSP auction

GSP: Truthful bidding?

AGT-MIRI Sponsored search

slide-56
SLIDE 56

Keyword Auctions Basic model GSP auction

GSP: Truthful bidding?

Consider a simple scenario.

AGT-MIRI Sponsored search

slide-57
SLIDE 57

Keyword Auctions Basic model GSP auction

GSP: Truthful bidding?

Consider a simple scenario. Two slots positions, with α1 = 200 and α2 = 100. All γi = 1

AGT-MIRI Sponsored search

slide-58
SLIDE 58

Keyword Auctions Basic model GSP auction

GSP: Truthful bidding?

Consider a simple scenario. Two slots positions, with α1 = 200 and α2 = 100. All γi = 1 Consider a bidder with value 10

AGT-MIRI Sponsored search

slide-59
SLIDE 59

Keyword Auctions Basic model GSP auction

GSP: Truthful bidding?

Consider a simple scenario. Two slots positions, with α1 = 200 and α2 = 100. All γi = 1 Consider a bidder with value 10 Facing competing bids of 4 and 8.

AGT-MIRI Sponsored search

slide-60
SLIDE 60

Keyword Auctions Basic model GSP auction

GSP: Truthful bidding?

Consider a simple scenario. Two slots positions, with α1 = 200 and α2 = 100. All γi = 1 Consider a bidder with value 10 Facing competing bids of 4 and 8.

Bidding 10 wins top slot, pay 8: profit 200 2 = 400. Bidding 5 wins next slot, pay 4: profit 100 6 = 600.

AGT-MIRI Sponsored search

slide-61
SLIDE 61

Keyword Auctions Basic model GSP auction

GSP: Truthful bidding?

Consider a simple scenario. Two slots positions, with α1 = 200 and α2 = 100. All γi = 1 Consider a bidder with value 10 Facing competing bids of 4 and 8.

Bidding 10 wins top slot, pay 8: profit 200 2 = 400. Bidding 5 wins next slot, pay 4: profit 100 6 = 600.

If competing bids are 6 and 8, better to bid 10. . .

AGT-MIRI Sponsored search

slide-62
SLIDE 62

Keyword Auctions Basic model GSP auction

GSP: Truthful bidding?

Consider a simple scenario. Two slots positions, with α1 = 200 and α2 = 100. All γi = 1 Consider a bidder with value 10 Facing competing bids of 4 and 8.

Bidding 10 wins top slot, pay 8: profit 200 2 = 400. Bidding 5 wins next slot, pay 4: profit 100 6 = 600.

If competing bids are 6 and 8, better to bid 10. . . It is not a dominant strategy to bid “truthfully”

AGT-MIRI Sponsored search

slide-63
SLIDE 63

Keyword Auctions Basic model GSP auction

GSP: Nash equilibrium

AGT-MIRI Sponsored search

slide-64
SLIDE 64

Keyword Auctions Basic model GSP auction

GSP: Nash equilibrium

A NE is a profile of bids b1 ≥ b2 ≥, . . . , ≥ bn such that,if π is the allocation of the GSP,

AGT-MIRI Sponsored search

slide-65
SLIDE 65

Keyword Auctions Basic model GSP auction

GSP: Nash equilibrium

A NE is a profile of bids b1 ≥ b2 ≥, . . . , ≥ bn such that,if π is the allocation of the GSP, for any player j, for k < j, αj(γπ(j)vπ(j) − γπ(j+1)bπ(j+1)) ≥ αk(γπ(j)vπ(j) − γπ(k)bπ(k)) and, for k ≥ j, αj(γπ(j)vπ(j)−γπ(j+1)bπ(j+1)) ≥ αk(γπ(j)vπ(j)−γπ(k+1)bπ(k+1))

AGT-MIRI Sponsored search

slide-66
SLIDE 66

Keyword Auctions Basic model GSP auction

GSP: Nash equilibrium

A NE is a profile of bids b1 ≥ b2 ≥, . . . , ≥ bn such that,if π is the allocation of the GSP, for any player j, for k < j, αj(γπ(j)vπ(j) − γπ(j+1)bπ(j+1)) ≥ αk(γπ(j)vπ(j) − γπ(k)bπ(k)) and, for k ≥ j, αj(γπ(j)vπ(j)−γπ(j+1)bπ(j+1)) ≥ αk(γπ(j)vπ(j)−γπ(k+1)bπ(k+1)) A player decreasing his bid can acquire a lower slot paying the price the player in this slot is paying.

AGT-MIRI Sponsored search

slide-67
SLIDE 67

Keyword Auctions Basic model GSP auction

GSP: Nash equilibrium

A NE is a profile of bids b1 ≥ b2 ≥, . . . , ≥ bn such that,if π is the allocation of the GSP, for any player j, for k < j, αj(γπ(j)vπ(j) − γπ(j+1)bπ(j+1)) ≥ αk(γπ(j)vπ(j) − γπ(k)bπ(k)) and, for k ≥ j, αj(γπ(j)vπ(j)−γπ(j+1)bπ(j+1)) ≥ αk(γπ(j)vπ(j)−γπ(k+1)bπ(k+1)) A player decreasing his bid can acquire a lower slot paying the price the player in this slot is paying. A player increasing his bid can only acquire a higher slot paying not the price the player in this slot is paying but its bid.

AGT-MIRI Sponsored search

slide-68
SLIDE 68

Keyword Auctions Basic model GSP auction

GSP: Nash equilibrium

AGT-MIRI Sponsored search

slide-69
SLIDE 69

Keyword Auctions Basic model GSP auction

GSP: Nash equilibrium

The NE equation have in general more than one solution, so there are many NE.

AGT-MIRI Sponsored search

slide-70
SLIDE 70

Keyword Auctions Basic model GSP auction

GSP: Nash equilibrium

The NE equation have in general more than one solution, so there are many NE. We have also a social welfare. PoS? PoA?

AGT-MIRI Sponsored search

slide-71
SLIDE 71

Keyword Auctions Basic model GSP auction

GSP: Nash equilibrium

The NE equation have in general more than one solution, so there are many NE. We have also a social welfare. PoS? PoA? Among the NE are there some with nice properties?

AGT-MIRI Sponsored search

slide-72
SLIDE 72

Keyword Auctions Basic model GSP auction

GSP: Nash equilibrium

The NE equation have in general more than one solution, so there are many NE. We have also a social welfare. PoS? PoA? Among the NE are there some with nice properties? Is there an efficient NE?

AGT-MIRI Sponsored search

slide-73
SLIDE 73

Keyword Auctions Basic model GSP auction

GSP: Envy-free equilibrium

AGT-MIRI Sponsored search

slide-74
SLIDE 74

Keyword Auctions Basic model GSP auction

GSP: Envy-free equilibrium

Definition Given a GSP with n players defined by click-through-rates α1 ≥ α2 ≥ · · · ≥ αn, quality scores γ1 ≥ γ2 ≥ · · · ≥ γn and valuations v1, . . . , vn. A bid vector b is an envy-free equilibrium if, for any pair j, k of players, player j would not prefer player k’s allocation and payments rather than their own. Formally αj(γπ(j)vπ(j) − γπ(j+1)bπ(j+1)) ≥ αk(γπ(j)vπ(j) − γπ(k+1)bπ(k+1)) where π(j) is the allocation of the GSP auction.

AGT-MIRI Sponsored search

slide-75
SLIDE 75

Keyword Auctions Basic model GSP auction

GSP: Envy-free equilibrium

AGT-MIRI Sponsored search

slide-76
SLIDE 76

Keyword Auctions Basic model GSP auction

GSP: Envy-free equilibrium

Every envy-free equilibrium is a NE

AGT-MIRI Sponsored search

slide-77
SLIDE 77

Keyword Auctions Basic model GSP auction

GSP: Envy-free equilibrium

Every envy-free equilibrium is a NE Conditions in k ≥ j are the same and, for k < j te conditions are stricter.

AGT-MIRI Sponsored search

slide-78
SLIDE 78

Keyword Auctions Basic model GSP auction

GSP: Envy-free equilibrium

Every envy-free equilibrium is a NE Conditions in k ≥ j are the same and, for k < j te conditions are stricter. Every envy-free equilibrium is efficient (exercise)

AGT-MIRI Sponsored search

slide-79
SLIDE 79

Keyword Auctions Basic model GSP auction

GSP: Envy-free equilibrium

Every envy-free equilibrium is a NE Conditions in k ≥ j are the same and, for k < j te conditions are stricter. Every envy-free equilibrium is efficient (exercise) Are there envy-free equilibria?

AGT-MIRI Sponsored search

slide-80
SLIDE 80

Keyword Auctions Basic model GSP auction

GSP: Envy-free equilibrium

Every envy-free equilibrium is a NE Conditions in k ≥ j are the same and, for k < j te conditions are stricter. Every envy-free equilibrium is efficient (exercise) Are there envy-free equilibria? Sort bidders so that γ1v1 ≥ · · · ≥ γnvn. Consider the bid vector b,

AGT-MIRI Sponsored search

slide-81
SLIDE 81

Keyword Auctions Basic model GSP auction

GSP: Envy-free equilibrium

Every envy-free equilibrium is a NE Conditions in k ≥ j are the same and, for k < j te conditions are stricter. Every envy-free equilibrium is efficient (exercise) Are there envy-free equilibria? Sort bidders so that γ1v1 ≥ · · · ≥ γnvn. Consider the bid vector b, b1 = v1 and, for i = j

AGT-MIRI Sponsored search

slide-82
SLIDE 82

Keyword Auctions Basic model GSP auction

GSP: Envy-free equilibrium

Every envy-free equilibrium is a NE Conditions in k ≥ j are the same and, for k < j te conditions are stricter. Every envy-free equilibrium is efficient (exercise) Are there envy-free equilibria? Sort bidders so that γ1v1 ≥ · · · ≥ γnvn. Consider the bid vector b, b1 = v1 and, for i = j bi = 1 αi−1γi  

j=i

n(αj−1 − αj)γjvj  

AGT-MIRI Sponsored search

slide-83
SLIDE 83

Keyword Auctions Basic model GSP auction

GSP: Envy-free equilibrium

Every envy-free equilibrium is a NE Conditions in k ≥ j are the same and, for k < j te conditions are stricter. Every envy-free equilibrium is efficient (exercise) Are there envy-free equilibria? Sort bidders so that γ1v1 ≥ · · · ≥ γnvn. Consider the bid vector b, b1 = v1 and, for i = j bi = 1 αi−1γi  

j=i

n(αj−1 − αj)γjvj   This is a envy-free equilibrium!

AGT-MIRI Sponsored search

slide-84
SLIDE 84

Keyword Auctions Basic model GSP auction

NE efficiency?

AGT-MIRI Sponsored search

slide-85
SLIDE 85

Keyword Auctions Basic model GSP auction

NE efficiency?

AGT-MIRI Sponsored search

slide-86
SLIDE 86

Keyword Auctions Basic model GSP auction

NE efficiency?

Are alll NE efficent?

AGT-MIRI Sponsored search

slide-87
SLIDE 87

Keyword Auctions Basic model GSP auction

NE efficiency?

Are alll NE efficent? Take 1 2 α 1 1/2 v 1 1/2 γ 1 1 b 1/2

AGT-MIRI Sponsored search

slide-88
SLIDE 88

Keyword Auctions Basic model GSP auction

NE efficiency?

Are alll NE efficent? Take 1 2 α 1 1/2 v 1 1/2 γ 1 1 b 1/2 b is a NE and its efficiency is 1 1

2 + 1 21 = 1.

AGT-MIRI Sponsored search

slide-89
SLIDE 89

Keyword Auctions Basic model GSP auction

NE efficiency?

Are alll NE efficent? Take 1 2 α 1 1/2 v 1 1/2 γ 1 1 b 1/2 b is a NE and its efficiency is 1 1

2 + 1 21 = 1.

The optimal allocation has efficiency 1 1 + 1

2 1 2 = 5/4

AGT-MIRI Sponsored search

slide-90
SLIDE 90

Keyword Auctions Basic model GSP auction

NE efficiency?

Are alll NE efficent? Take 1 2 α 1 1/2 v 1 1/2 γ 1 1 b 1/2 b is a NE and its efficiency is 1 1

2 + 1 21 = 1.

The optimal allocation has efficiency 1 1 + 1

2 1 2 = 5/4

PoA?

AGT-MIRI Sponsored search

slide-91
SLIDE 91

Keyword Auctions Basic model GSP auction

PoA

AGT-MIRI Sponsored search

slide-92
SLIDE 92

Keyword Auctions Basic model GSP auction

PoA

In the full information setting the quality factors γ are fixed and common knowledge.

AGT-MIRI Sponsored search

slide-93
SLIDE 93

Keyword Auctions Basic model GSP auction

PoA

In the full information setting the quality factors γ are fixed and common knowledge. Theorem The (pure) PoA of GSP in the full information setting is at most the golden ratio 1

2(1 +

√ 5) ≈ 1.618

AGT-MIRI Sponsored search

slide-94
SLIDE 94

Keyword Auctions Basic model GSP auction

PoA

In the full information setting the quality factors γ are fixed and common knowledge. Theorem The (pure) PoA of GSP in the full information setting is at most the golden ratio 1

2(1 +

√ 5) ≈ 1.618 and at least 1.282 .

AGT-MIRI Sponsored search

slide-95
SLIDE 95

Keyword Auctions Basic model GSP auction

Design directions

AGT-MIRI Sponsored search

slide-96
SLIDE 96

Keyword Auctions Basic model GSP auction

Design directions

Main focus in the extension of keyword auctions to other settings.

AGT-MIRI Sponsored search

slide-97
SLIDE 97

Keyword Auctions Basic model GSP auction

Design directions

Main focus in the extension of keyword auctions to other settings. Goal: Design mechanisms that verify properties. For example:

AGT-MIRI Sponsored search

slide-98
SLIDE 98

Keyword Auctions Basic model GSP auction

Design directions

Main focus in the extension of keyword auctions to other settings. Goal: Design mechanisms that verify properties. For example: Individual Rationality: Each player has net non-negative utility from participating in the auction, i.e., ui ≥ 0.

AGT-MIRI Sponsored search

slide-99
SLIDE 99

Keyword Auctions Basic model GSP auction

Design directions

Main focus in the extension of keyword auctions to other settings. Goal: Design mechanisms that verify properties. For example: Individual Rationality: Each player has net non-negative utility from participating in the auction, i.e., ui ≥ 0. Incentive compatibility (a.k.a. truthfulness): It is a dominant strategy for each player to participate in the auction and report their true value.

AGT-MIRI Sponsored search

slide-100
SLIDE 100

Keyword Auctions Basic model GSP auction

Design directions

Main focus in the extension of keyword auctions to other settings. Goal: Design mechanisms that verify properties. For example: Individual Rationality: Each player has net non-negative utility from participating in the auction, i.e., ui ≥ 0. Incentive compatibility (a.k.a. truthfulness): It is a dominant strategy for each player to participate in the auction and report their true value. Pareto-optimality: An allocation π and payments p is Pareto-optimal if and only if there is no alternative allocation and payments where all players’ utilities and the revenue of the auctioneer do not decrease, and at least one of them increases.

AGT-MIRI Sponsored search