1998 Feb 01

Aptex SelectCast
David M. Raab
DM News
February, 1998

Software developers have spent much energy over the past few years trying to automate the production of statistical models. With products like ASA ModelMAX, Group One Model1, Trajecta dbProphet, Urban Science GainSmart and the forthcoming SAS Enterprise Miner, one might argue that they have largely succeeded–or, at least, have taken the process about as far it can go. Even the latest systems still need a skilled user to produce the best results.

But while vendors have pursued more efficient development of individual models, marketers’ needs have grown more complex. Faced with more demanding customers, expanding product lines and, above all, the Internet, they need to choose which among hundreds or even thousands of offers to present to each customer. And they need to do it quickly–sometimes while the customer waits for a response. And they need to do it for thousands or millions of people with unique histories. And they need to incorporate new offers and customers without advance testing.

This is not a set of tasks that are suited to conventional modeling techniques, which mostly identify the best prospects for one or a few products. Even if production of the individual models is automated, these techniques do not provide a natural method to compare the value of different offers to the same person. This takes a different approach entirely.

SelectCast (Aptex, 619-623-0554, www.aptex.com) is designed to handle the challenge of choosing among multiple offers for different individuals. The system is targeted specifically at Internet advertising, both in choosing which “banner” ads to display as an individual navigates a Web site and in deciding which products to offer as someone browses an online catalog. A sister product, SelectResponse, uses the same underlying technology to automate responses to other types of interactions, such as customer service requests.

The core technology involves a technique called “context vector mining”, which analyzes text to determine which words are frequently used together and then assigns each word a measure, called a vector, that is close for related words. The vectors for individual words can be combined to create a combined vector for a document or set of documents. Vectors are positioned in a multidimensional space, which allows calculations to show how “close” one vector is to another. A typical SelectCast analysis might use several hundred vectors.

The actual process of vector creation is proprietary, complex, and of little interest to end-users. The system uses a neural network algorithm and is derived from work done by Aptex’s corporate parent, HNC, a major developer of neural network systems.

In practice, SelectCast works by calculating the position of each ad and each individual viewer and then selecting the ad that is closest to the viewer being evaluated. For banner ads, the position is calculated by analyzing the contents of the Web page the ad jumps to. For individuals, the position is determined by the positions of the pages the individual has already viewed.

The positions are further adjusted each time a viewer is presented with the option to select a given ad: they move closer together if the ad is selected, and farther apart if it is not. This lets the system learn the most effective position for a given ad and adjusts the position automatically as conditions change. Since a flurry of atypical events could pull an ad away from its best position, the system also tracks the response rate to banner presentations and will restore earlier settings if the rate declines over time.

A different approach is needed for individual viewers, whose locations truly do shift as their objectives change. SelectCast lets the marketer control how heavily it weighs the most recent transactions and how quickly it “forgets” the older ones. In a shopping situation, the system can also apply different weights to items the user merely browses, puts in a virtual “shopping cart”, and actually purchases.

Vectors for both advertisements and individual viewers are stored in a central database. The viewer profiles are linked to a “cookie” or digital certificate on the viewer’s own computer, so individual identities are not stored centrally.

Of course, in the real world marketers have other objectives beyond simply getting the highest possible response rate. For advertising, SelectCast can manage an inventory of advertisements with start and stop dates, a guaranteed number of impressions, and how many other ads must intervene before the same ad is shown to someone again.

For merchandise offers or ads sold on a cost-per-click-through basis, the system can also accept different values for different ads and attempt to maximize revenue rather than response rates. However, it does not manage the quality of the people it presents with an ad–say, by not presenting luxury auto ads to young boys who might be interested but are unlikely to purchase. SelectCast does report on the inferred age and gender of the people who view different ads, although the inferences are based on generic behavior profiles derived from a combination of sites. Other reports show response rates based on behavior categories that are site-specific.

SelectCast recognizes that who responds to an ad is determined in large part by who is shown it, and automatically exposes a sample of lower-ranking candidates to avoid self-fulfilling prophecies. It also supports comparison with alternative techniques through a random selection function that lets the marketer exclude a fraction of the audience from SelectCast operations. However, the system does not allow testing of alternative vector definitions or decision rules within SelectCast itself.

SelectCast was introduced in 1996 and currently has eight installations for advertising and six installations for merchandise applications. It governs banner selection for many major Web advertisers, including Excite, InfoSeek, DoubleClick, InfoSpace and MatchLogic. It has handled databases over 25 million users and volumes over 300 transactions per minute. The system runs on a Sun server and can integrate with NetGravity and other ad server software on both Sun and Windows NT platforms. Prices are upwards of $100,000, based on a license fee plus site ad revenues.

* * *

David M. Raab is a Principal at Raab Associates Inc., a consultancy specializing in marketing technology and analytics. He can be reached at draab@raabassociates.com.

Leave a Reply

You must be logged in to post a comment.