Presentation is loading. Please wait.

Presentation is loading. Please wait.

Ad Exchanges: Research Issues S. Muthukrishnan Google Inc. Presented by Tova Wiener, CS286r 11/16/2009.

Similar presentations


Presentation on theme: "Ad Exchanges: Research Issues S. Muthukrishnan Google Inc. Presented by Tova Wiener, CS286r 11/16/2009."— Presentation transcript:

1 Ad Exchanges: Research Issues S. Muthukrishnan Google Inc. Presented by Tova Wiener, CS286r 11/16/2009

2 Framework Advertising on the Internet involves three parties:  Users  Publishers  Advertisers An Ad Exchange brings sellers (publishers) and buyers (advertisers) together to a single marketplace Generally, contracts are sold in advance; however, publishers who have ads that are not sold can send their slots to the exchange to sell. Websites customize ads based on their knowledge about the current user. Ad exchanges ease the burden on website publishers by accepting bids from many networks and selecting the revenue maximizing bid for the publisher.

3 Sequence of Events User u visits webpage w of publisher P(w) Publisher P(w) contacts the exchange E with (w,P(u),p) The exchange E contacts the ad networks with (E(w),E(u),p) w (w, P(u), p ) E a1a1 aiai amam (E(w), E(u), p ) Website ExchangeNetworksAdvertisers

4 Sequence of Events Each ad network returns (b i,d i ) on behalf of its advertisers. Exchange E determines a winner i* for the slot and a price c i*. P(w) serves the webpage w with the ad to the user u. w (w, P(u), p ) E (i*,c i* ) a1a1 aiai amam (E(w), E(u), p ) (b i,d i ) Website ExchangeNetworksAdvertisers

5 Observations on the Model The units traded are impressions  Display ads focus on marketing, not actions.  Difficult to collect accurate click-through data. Who determines what the advertisers know about the website? How is “fit” between advertisers and websites measured? Computation must happen in microseconds. Can each website publisher only be involved with a single exchange?  The publisher must accept the price returned by the Exchange, and trust the Exchange to do content monitoring The advertisers must trust the networks to bid to maximize their individual values. How does this differ from the financial auction model?  Heterogeneity and perishable goods

6 Basic Auction at the Exchange Assume there is a single slot being auctioned, and each ad network submits a bid for its advertisers, and that the Exchange uses a second price auction. Define book value to be the second largest value of all bidders Problem 1: Assuming p is exogenous and assuming the advertisers reveal their bids truthfully to the networks, is there a possibly truthful auction at the exchange that will extract a large fraction of the book value?

7 Basic Auction at the Exchange Why would a network ever reveal their second highest bidder?  More advertising spots on each webpage  Auctions run in the “long-run” where the highest bidder has budget constraints  Other possible incentives? Can we reasonably bound the expected difference between the second highest bidder of one network and the highest bidder of another network?  If there are enough networks, and each is big enough, will demand really be that widely distributed?

8 Auction and Bidding by Ad Networks Problem 2 Assuming p is exogenous, and the exchange runs a second price auction with reserve price r > p, ie., E(p) = r, and advertisers are captive, that is, remain with their choice of ad network throughout, what is a revenue optimal mechanism for an ad network? This refers to the mechanism that the network will use to extract bids from the advertisers. Why is this auction different than standard auctions in which we can use VCG?

9 Auction and Bidding by Ad Networks Why is this auction non-standard?  The networks may not know the distribution of bids beforehand.  They are selling a contingent good: there is some probability ® (b), when the network submits some bid b that they will win, based on the bids of other networks. How will strategies change as networks learn about the types of advertisers in other networks? How should a network bid within the exchange to maximize its own revenue? This is the of the Ghosh et al. paper.  How should networks charge their bidders?  What would be optimal if they know that the Exchange is running GSP or VCG?

10 Auctioning with Heterogeneous Valuations The optimal revenue for an auction is R * = max i v i, where v i, is bidder i’s valuation for the impression. How close can we get to generating revenue R*? If values are truly heterogeneous, then VCG and GSP (ie, looking at second prices) will not work. The problem has been solved for the case where a prior distribution over the valuations is known. No truthful mechanism exists with (roughly) expected revenue £ (R*/log R*), but a randomized algorithm has been shown with revenue close to that. Problem 3 Design a non-truthful mechanism for prior- free auction of a single slot with near-optimal revenue, but with good equilibrium properties.

11 Auctioning with Heterogeneous Valuations By good equilibrium properties, they mean a Complete Information Nash Equilibrium.  The ad networks have priors about the other networks bidding behavior, but the situation is prior free because the Exchange does not have these priors. They leave the idea of truthfulness in order to increase revenue: what are the properties of this tradeoff. Quasi-proportional Allocations: the ith bidder is selected with probability f(b i )/  i f(b i ). If f(x) = x this is simply the proportional allocation rule. They hypothesize that quasi-proportional allocations are the right thing to do in this case. Why?

12 Auctioning with Heterogeneous Valuations What if we relax the assumption that the networks have priors about the bidding practices of other networks? Problem 4 Design (even non-truthful) mechanisms for prior-free bidding of ad networks in AdX, with good equilibrium properties and (near-)optimal revenue. How does being an incomplete information setting change things? Again, they need to relax truthfulness to allow for prior-free bidding. What is the revenue of the ad networks in the first place? Do they charge a constant reserve price, or take a percentage or what is paid to the exchange?

13 AdX integrity The networks and advertisers must trust AdX to participate Problem 8 Design a cryptographically sound real-time auction protocol so that any participating party in AdX can verify that (a) all communication, accounting and computations were performed correctly, and (b) auction was closed envelope, that is, no bidder sees others’ bids prior to the auction. This has to work for repeated auction of impressions in AdX where some information is revealed between impressions. While such cryptographic protocols have been explored, they need to be fast. Also, must be expressed through a clear model so that networks can prove the strong cryptographic properties of various protocols.

14 Callout Optimization How should the Exchange get bids from the ad networks without consuming too much bandwidth? Consider the approach where E makes http calls to the networks servers and waits for the networks to reply with bids. Does the exchange need to communicate with each network for every ad?  Networks can inform exchanges about the general types of impressions that their customers would be interested in, and then the exchange can only query certain networks.  This implies that the networks cover different parts of the advertising market. How much competition can we expect on each sector, as opposed to how much cooperation between networks to cover the whole market space?

15 Callout Optimization Problem 5 Each ad network i has bandwidth budget B i. Say E has bandwidth budget of B. Design an online algorithm for E that for each incoming call (w j, u j, ½ j ), chooses a subset S j µ S (E(wj ),E(uj ), ½ j) of networks to call such that no ad network i gets more than B i calls per second, E make fewer than B calls per second, and optimizes the expected:  number of bids, ie, number of nonempty (b i (j), d i (j))’s received at E, or  efficiency  j max i b i (j), or  sales revenue  j max i|bi(j)  maxi bi(j) b i (j), or  profit for E. Moreover, we need to learn the probability that each network will bid a non-empty, or relevant, value. What would the optimization problems look like for each of these cases?

16 Conclusions The research topics presented here are important for the future growth of the market  Game Theory of Advertisers: Advertisers may go to multiple networks, or choose networks strategically. How does this affect exchange dynamics?  Ad Quality: We require a quality metric to price incentives endogenously. A proposal is to generate a suitable Markov model for users that will capture even the long term impact of ad impressions. (That seems like a pretty hard problem, how would we categorize users?) On a higher level, how can we think about this “multi-level” auction set-up? Are there other instances where this kind of contingency would be a factor? Could one imagine three or four level auctions?

17 Publisher Optimization and Strategies Now we deal with how the website publisher should choose to interact with AdX. Why would Google be trying to solve this problem? Problem 6 Given models for impressions inventory (w,u), models for bids (b i*,d i* ) from E, models of ad sales and prices through other channels, design an algorithm that on each impression (a) decides whether to go to AdX, (b) chooses disclosed or undisclosed inventory at AdX, and (c) selects min price p, in order to optimize the expected overall (long term) revenue.


Download ppt "Ad Exchanges: Research Issues S. Muthukrishnan Google Inc. Presented by Tova Wiener, CS286r 11/16/2009."

Similar presentations


Ads by Google