Friday, October 5, 2012

How Can Research Optimize the Advertising World?

Ana Radovanovic of Google gave this talk on optimizing advertising.  She started off by giving some background on the advertising ecosystem at Google.

The goal of Google is to organize the world's info and make it universally accessible and useful.  An example of this is their ad system: ads are displayed on the right side of the search page.  Why put ads on the side?  It's actually the result of common sense thinking: hundreds of millions of users every day, in 190 countries and 41 languages click on ads, so it makes sense to offer relevant ads there.  Google also provides the ability to do contextual advertising on any webpage:

The google ad network ecosystem is set up to give users ony relevant ads, provide publishers with maximal revenue per ad shonw, and allow advertisers to maximize the return on money spent on advertising.  Note that it's easy to trade off the interests of users, advertisers, and publishers, but through clever technologies it's also possible improve the overall efficiency without trading off interests.  Such technology can make the network extremely efficient.

Ana also talked about targeting criteria: the advertiser's belief of where the ad is the most profitable.  This criteria includes:
  • keywords
  • networks
  • industry verticals
  • sites
  • language
  • geo targets
  • ad schedule
In machine learning terms, this is the 'prior information'.

Optimization: how can ads be targeted efficiently?
One important aspect of optimization is click-through rate prediction: or trying to learn what ads will result in a click by a user.  If we know in advance how often this will happen, we can estimate the ad ranking function.  There are some challenges though, i.e. getting really real-time prediction, doing online learning, and executing a parallel learning algorithm.

Google's ad auction is a sort of middle ground between the three interested parties (users, publishers, advertisers).  But there are some machine learning problems related to it:
  • Budget optimization: predicitng the rate at which to show the ad throughout the day to spend the budget in the most efficient way
  • Spam detection: detecting spammy impressions/clicks
  • Smart pricing: how to adjust the price of an ad to equalize ROI between sites that perform differently
Ana showed some of the formulas used to determine how to maximize an advertiser's profit (I won't show those here as her slides will be posted on the wiki). Once ads are displayed, they are also ranked by relevance to benefit the user.

One interesting idea Ana mentioned is that some advertisers aren't buying ads to actually sell products (she used Coca Cola as an example).  This is brand marketing, and the idea is just to get the brand in your mind, not necessarily for you to make an immediate purchase.  How should advertising be priced/displayed to satisfy these advertisers?

No comments: