CATS-Feb-21-2014
Title
Revenue Monotone Mechanisms for Online Advertising
Speaker
Reza Khani, University of Maryland
Abstract
Online advertising is an essential part of the Internet and the main source of revenue for many web-centric firms such as search engines, social networks, and online publishers. A key component of online advertising is the auction mechanism which selects and prices the set of winning ads.
This work is inspired by one of the biggest practical drawbacks of the widely popular Vickrey-Clarke-Groves (VCG) mechanism, which is the unique incentive-compatible mechanism that maximizes social welfare. It is known that VCG lacks a desired property of revenue monotonicity - a natural notion which states that the revenue of a mechanism shouldn't go down as the number of bidders increase or if the bidders increase their bids. Most firms which depend on online advertising revenue have a large sales team to attract more bidders on their inventory as the general belief is that more bidders will increase competition, and hence revenue. However, the lack of revenue monotonicity of VCG conflicts with this general belief and can be strategically confusing for the firm's business.
In this work, we seek incentive-compatible mechanisms that are revenue-monotone. This natural property comes at the expense of social welfare - one can show that it is not possible to get incentive-compatibility, revenue-monotonicity, and optimal social welfare simultaneously. In light of this, we introduce the notion of Price of Revenue Monotonicity (\porm) to capture the loss in social welfare of a revenue-monotone mechanism.
We further study revenue-monotonicity for two important online advertising scenarios. First one is the text vs image ad auction where in an ad slot, one can either show a single image ad or a few text ads. Second one is the video-pod auction where we have a video advertising slot of $k$ seconds which can be filled with multiple video ads. For the image-text auction, we give a mechanism that satisfy both RM and IC and achieve \porm{} of $\sum_{i=1}^k \frac{1}{i} \approx \ln k$. We also show that the \porm{} of our mechanism is the best possible by proving a matching lower bound of $\sum_{i=1}^k \frac{1}{i}$ on the \porm{} of any deterministic mechanism under some mild assumptions. For the video-pod auction, we give a mechanism that achieves a \porm{} of $(\lfloor \log k \rfloor + 1) \cdot (2 + \ln k)$.