Two Different Paths to PersonalizationPersonalization long ago became such a popular buzzword that it has lost any specific meaning. This means that describing any piece of personalization software requires a preliminary tour of the marketplace to define where that software fits in. So here goes.
In the broadest sense, personalization can be applied to any system that treats different people differently. This includes operational systems that deliver the personalized content, such as call centers or Web sites, as well as selection systems that choose the content for each individual. Most of today's operational systems include their own selection mechanisms, but many other selection systems exist independently.
Selection systems fall into two major categories. Rule-based systems rely on specific if/then logic to handle each situation. This approach is precise and easy to understand, but it usually requires considerable labor and relies heavily on the user's expertise. Automated rule-generation systems do exist, but even these need a fair amount of human oversight.
Other selection systems use some sort of scoring or modeling. These may employ conventional statistical techniques such as regression, various breeds of neural network models, Bayesian networks, clustering methods such as collaborative filtering or other approaches.
For better or worse, most marketers have little interest in the technical details. But marketers do need to make at least one important distinction: between systems that require separate management of each offer and those that incorporate new offers automatically. Systems in the first group require the user to define the characteristics of an offer, when it can be presented, how it relates to other offers and other features. This labor places a practical limit on the total number of offers that users can load into such a system. Situations with hundreds of thousands or millions of offers - such as recommending books, music or newspaper articles - need a fully automated approach.
Automated offer management systems use either of two methods. The first method, generally called collaborative filtering, makes no direct attempt at assigning attributes to the offers themselves; instead, it looks for customers who behave similarly and identifies offers that those customers select. The second approach classifies the offers; it recommends the offer most similar to whatever the customer selected last. The first method makes the most sense when there are large numbers of customers and it is difficult to discern the objective differences among offers. While easy to deploy, it is often insensitive to shifts or peculiarities in an individual customer's interests. The second method is more sensitive to individual behavior but requires sophisticated techniques to identify offer attributes. What both methods share is the ability to find patterns and similarities without direct human intervention.
Personal Sales E.ssistant and Personal Information E.ssistant (Dynaptics, 408/918-2400, www.dynaptics.com) are from the same company and use Bayesian scoring techniques. But they occupy opposite poles of the personalization spectrum. Sales E.ssistant requires individual offer definition, while Information E.ssistant automatically identifies themes based on offer text.
Sales E.ssistant is designed primarily to display offers on personalized Web sites. Each offer is defined to the system with an identifier, text and product ID. The user also specifies the location of the image to display with the offer and the destination, such as an order entry page, when the offer is selected. The user then defines where and when the offer may be displayed, based on a date range and a hierarchical map of the Web site. The user also can assign a weighting factor that will give the offer a higher or lower priority based purely on the system-generated scores. This lets marketers test new offers quickly and force specified offers to appear in particular situations. Users also control how often the same offer may be repeated within a Web session and how long to wait before reshowing an accepted offer, but these rules are set at the system level and apply equally to all offers.
Sales E.ssistant is integrated with a Web site through HTML tags that are activated when a visitor requests a given Web page. These tags send a message to Dynaptics that identifies the visitor, page and page category, with the option of several other user-defined attributes. Dynaptics uses these messages to track the path of each visitor through the Web site and returns the highest-scoring offer subject to eligibility rules and user-specified weighting. Selections also can take into account previous behavior and other customer data, although this may require custom integration.
Dynaptics scores are determined by acceptance rates of previous customers who visited the same sequence of pages as the current visitor. This lets the system respond to current behavior, but it means scores may be unreliable if only a few people have followed the same path previously. Dynaptics uses special statistical methods to deal with this and also lets the user set the number of pages in a sequence. The default is to track the last six page views. Scoring statistics are updated immediately after each offer is presented, so any change in overall customer behavior is reflected at once. Real-time reports show the acceptance rate by offer.
Unlike Sales E.ssistant, Information E.ssistant requires no setup for individual offers - which, in this case, are text documents such as newspaper articles. The system instead analyzes the contents of these documents by looking at the frequencies of individual words. After comparing these frequencies for several documents selected by a visitor, it can determine which other documents are most similar to those already seen. Because it finds relationships among the specific documents selected by each visitor, Information E.ssistant can identify common themes without using key words or any other predefined classifiers.
Both Sales E.ssistant and Information E.ssistant run on Solaris or Linux servers and can be operated inhouse or by the vendor. Prices are based on volume and begin at $75,000 each for an inhouse license. The systems were released in 2000 and have about a half-dozen pilot implementations.