Download presentation
Presentation is loading. Please wait.
Published byTrevor Taylor Modified over 9 years ago
1
Authors: David Robert Martin Thompson Kevin Leyton-Brown Presenters: Veselin Kulev John Lai Computational Analysis of Position Auctions
2
Motivation Many different models of ad auctions Each model is partially understood Multiple equilibria e.g. Locally-envy free equilbria Hard to theorize about full set of equilibria Use computational techniques to fill in the gap
3
Outline Different auction types, preference types Action graph games Experimental setup Experimental results Discussion
4
Auction Types Generalized First Price (GFP) i th highest bid is allocated slot payment is exactly the submitted bid Unweighted Generalized Second Price (uGSP) i th highest bid is allocated slot i payment is the (i+1) st highest bid Weighted Generalized Second Price (wGSP) each bid b j is multiplied by a bidder-specific weight w j order bids by b j * w j = effective bid for j i th highest effective bid is allocated slot i (call this agent k) payment is the (i + 1) st effective bid / w k
5
Preference Types Two Dimensions to Vary CTR: click through rate model Value: how much the user values a click Edelman et. al. (EOS) CTR: decreasing in position, same across bidders Value: same value for all clicks, regardless of position Varian (V) CTR: separable into position-specific and bidder-specific components; ctr(pos i, bidder j) = ctr(i) * score(j) Value: same as EOS (constant for all clicks)
6
Preference Types (cont.) Blumrosen et. al. (BHN) CTR: same as V (decreasing, bidder-specific but separable) Value: value per click increasing in rank; higher positions are valued more highly Benisch et. al (BSS) CTR: same as EOS (decreasing, bidder-independent) Value: single peaked in position; strictly decreasing from peak
7
Preference Types Summary CTR Independent of Bidder CTR is separable ( ctr(p, b) = ctr(p) * qual(b)) Value is Independent of Position EOSV Value Increases with Position ?BHN Value is Single PeakedBSS?
8
Formal Description
9
Questions EOS locally envy-free equilibria are efficient and VCG- revenue dominant how often does wGSP have efficient, VCG-revenue dominant? what happens in other equilibria? V any symmetric equilibrium (globally envy free) is efficient and VCG-revenue dominant how often does wGSP have efficient, VCG-revenue dominating equilibria?
10
Questions (cont.) BHN there are preferences where wGSP has no efficient NE how often does wGSP have no efficient NE? How much welfare is lost? BSS wGSP can be arbitrarily inefficient how often does wGSP have no efficient NE? How much social welfare is lost?
11
AGG Example Single Item First Price Auction Two bidders with values v1 = 4 and v2 = 6 Discretize and bounds bids B2=1B2=2B2=3B2=4B2=5B2=6 B1=1½(3)00000 B1=22½(2)0000 B1=311½(1)000 B1=4000000
12
AGG Example (cont.) AGG Representation b2 < 1b2=1b2 > 1 13½(3)0 b2 < 2b2=2b2 > 2 22½(2)0 b2 < 3b2=3b2 > 3 31½(1)0 AGG size not dependent on number of possible v2 bids or discretization
13
Action Graph Games normal form representation can be very large strict independencies Payoff for agent A is always independent of agents B’s action context-specific independencies Payoff for agent A is independent of action of agent B for some subset of actions for A and B e.g. First Price Auction: Payoff for agent A is independent of agent B’s action if agent B bids less than agent A
14
Why AGG? compact size (exponentially smaller) does not increase with more agents AGG structure can be leveraged computationally polynomial time algorithm (in the compact size) for computing expected utility of a strategy
15
Function Nodes nodes that are not actions, but are computed based on actions can be useful to decrease the in-degree of action nodes if each player affects the function nodes independently, can still find expected utility in polytime Example: GSP payoff depends on the number of bids higher than you, but not the identity of those bids
16
AGG Examples (cont.)
17
Experimental Setup Weakly dominated strategies removed Strategies where bidder bids higher than value Strategies where agent has bids j > i, where the allocation for the agent is the same for all bids of other agents Happens when weights are very different Impact on locally envy-free? Uniform Sampling
18
Experimental Results EOS Approximately efficient Did not beat VCG revenue even in best equilibria uGSP = wGSP more efficient than GFP Ambiguous revenue results (wGSP v. GFP) V Approximately efficient Did beat VCG revenue Dominated GFP, uGSP in efficiency Revenue only better than GFP, uGSP in medium
19
wGSP v. VCG Revenue Edelman only examines locally envy-free equilibria (other equilibria might exist) Bid interval may be empty Discretization Bids could be higher than bidder’s value
20
wGSP v. VCG (EOS)
21
Experimental Results BHN wGSP had frequent, complete failures of efficiency Discretized VCG also suffered from this wGSP had higher welfare than GFP, uGSP Ambiguous revenue results BSS Similar to BHN
22
Experimental Results Summary wGSP generally efficient Ambiguous revenue results (compared to VCG); lower for EOS, higher for V, ambiguous for BHN, BSS
23
Conclusion / Discussion wGSP has comparable performance to VCG Can leverage computation to help examine equilibria under different assumptions / mechanisms What do the “other” equilibria look like? Which equilibria are selected in practice? (hard to know)
24
Conclusion / Discussion How are weights computed? What happens if weights used by wGSP are not perfectly accurate? Analysis is for single keyword auctions; do bidders actually optimize at this level?
25
AGG Examples (cont.)
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.