Improved Ad Selection

Web advertising has gone a long way from its starts around 10 years ago. Yet, it is still easy to see that most of the ad content shown today brings little value to the users.

The next generation of ad selection technology should bring the relevance and the appropriateness of the ads closer to the other content served on the web. Some recent directions are exploring the intent of user queries for search advertising; using knowledge about the world wide web; retrieval and scoring mechanisms specifically targeted for advertising; combining relevance and click based feedback for ad selection; factoring the revenue in the ad selection; methods for rapid internationalization of ad selection, and modeling/machine learning techniques for better ad selection.

Behavioral targeting and social network advertising

Ads can also be selected based on past user behavior, as web searches, visited sites, browsing behavior, as well, as data known about the user: age, income, relationship to other users, etc. Such ad selection is called Behavioral Targeting. This is a new area of research and there is little published work how to effectively exploit such data in online advertising. It especially pertains to advertising on social networks that are becoming an increasing share of the overall Internet activity.

Advertiser experience

Another area of research is improving advertiser's experience and increasing the value that the advertiser gets from online advertising. The process of ad generation in today's systems is tedious: the advertisers need to craft their ads, select the right bid phrases and select bids such that they get the maximum value for their budgets. There are many ways how to improve the advertiser experience: better keyword selection; automatic ad generation; improved performance forecasting, etc.

Cost of serving and scalability

Today's major ad platforms are complex systems running on thousand of machines. To make the ad selection profitable, the cost of ad selection should not exceed the generated revenue. One of the research challenges is how to architect systems that would perform ad selection in orders of magnitude less computational effort than today. Some of the possible directions here are approximating/caching ad scores and ads; tiering architectures for ad selection, etc.

Forecasting user visits

One of the challenging problems in guaranteed-delivery display advertising, where advertisers can buy targeted user visits months in advance from online publishers, is the problem of forecasting user visits. For instance, if an advertiser wants to buy 100 million user visits by Males in California visiting Sports Pages in November 2010, then the publisher needs to have an accurate forecast of the number of user visits that satisfy that specification. However, forecasting user visits is a hard problem for two reasons. First, user visits are very high-dimensional, specifying the demographics, stated interests, and inferred interests of the user, the category and type of the page being visited, and other attributes such as IP location etc.; this problem is further compounded by the fact that advertisers can target any combination of the user visit attributes. Second, there is inherent uncertainty in user behavior that is hard to explicitly model.

Back to the Key Scientific Challenges Program