Too Much Information, Made Easier

I recently went to a well-known retailer's Web site and received messages offering new products there. Whether these were indeed recommendations, based on my Web surfing habits, or were just the specials of the day, I don't know. What I do know is that these recommendations met my needs by making offers that happened to match my interest. While there may be several ways to offer customers recommendations, all of them revolve around extracting the right customer "knowledge." Recently, there's been a sharp growth in the volume of information available to help make informed purchasing decisions. It's considerably more than anyone can possibly filter for a "correct" decision. Database and e-marketers face similar situations as they make marketing and merchandising decisions as to what next product or service to offer. Types of filter approaches yield accuracy differently Recommendation engines, or collaborative filtering solutions, analyze data by observing others with similar interests, and provide suggestions for a probable, reasonable next purchase. Consumers seeking advice on a digital camera may look for the "cumulative intelligence" of others with similar interests. Since there is so much data to collate, the marketer must selectively filter, and select the most suitable recommendation. A primary approach is via explicit ratings. This technique surveys a consumer and the response triggers a recommendation. This process is well defined and fairly accurate. Imagine, however, a user clicking his way through an e-commerce site and suddenly being presented with a survey. These few minutes completing the questionnaire can interfere with normal patterns of browsing. A more prudent method is to use implicit evaluations, where a rating is obtained indirectly. Such pervasive collaborative filtering methods are based on behavioral data that emerge from actual transaction and browsing behavior, and customer data sources. Most collaborative filtering solutions involve cluster analysis, where members of one cluster "look like" fellow members, and are dissimilar to members of a different cluster. Two primary implicit-filtering approaches are: o locating individuals similar to the user seeking a recommendation. This is referred to as "K-nearest neighbor" o using historical information to identify relationships where the purchase of one product or service will lead subsequently to the purchase of another The ROI of collaborative filtering To simulate results, we examined ownership of products from a single marketer for 42,206 individual customers. At random, we removed the most previous purchase from the customer's record. Ten recommendations were then made for each individual based on the balance of the transaction history. Our goal was to determine whether our "recommendation" included the product that was removed from each individual's record. In about 31 percent of cases, our recommendations included the removed product. Marketers may ask how good these suggestions are. If, for example, the marketer could potentially offer 500 different products, the likelihood of randomly assigning a product to a customer from an inventory of 10 products is 10/500, or 2 percent. Yet collaborative filtering technology generated a 31 percent hit, a 15-fold increase. While the approach discussed here is not new, many marketers are reluctant to get their feet wet with such data analytics. Some claim the investment may be too high. Others feel that a shopper knows what he or she wants, and does not need to be coerced. Those that fail to use these systems are losing out. The fact is these systems are working, and provide the marketer a clear advantage. There really is no viable way of negotiating all the possible choices that a consumer has. Recommendation engines provide a powerful means to assist the customer in making the right choice. After all, is this not what marketing is all about? About the Author Sam Koslowsky is vice president of modeling solutions for Harte-Hanks Inc., a direct and interactive marketing company, and a regular speaker on data modeling at Direct Marketing Association conferences. Contact him at sam_koslowsky@harte-hanks.com