Web Sites Learn To Make Smarter Suggestions

Reading Time: 3 min 
Permissions and PDF Download

With the pricking of the Internet bubble, online retailers are under more pressure than ever to earn their keep. As a result, many companies are looking to sophisticated merchandising tools, such as intelligent agents that recommend products, to build customer loyalty and sales.

Intelligent recommendation agents —automated recommendation systems that learn and improve over time — have traditionally had limitations. Most agents are based on collaborative filtering, which works by matching each target customer to a group of users with similar tastes — then using the group's choices to generate suggestions for the target customer. In order to come up with useful recommendations, collaborative filtering requires vast amounts of data, which many smaller businesses find difficult to obtain. Collaborative filtering relies on information from past users, so it's unlikely to recommend new or obscure products, even if they perfectly fit a customer's needs. And even under the best of circumstances, collaborative filtering sometimes generates recommendations that seem just plain bizarre.

Researchers in computer science and marketing are studying a variety of methods for improving the performance of recommendation agents. One possibility is to combine collaborative filtering (or other techniques based on customer communities) with an individual-based approach, which uses information from a single customer to figure out which product attributes she values most. In a recent computer simulation, individual-based agents initially performed relatively poorly. After about 40 trials, however, the simulated customers “purchased” products recommended by individual-based agents 80% of the time. By contrast, two common forms of collaborative filtering reached success rates of 50% to 60%.

Those results are reported in “Which Intelligent Agents Are Smarter? An Analysis of Relative Performance of Collaborative and Individual-Based Recommendation Agents,” by Manuel Aparicio, founder and chairman of Saffron Technology, based in Morrisville, North Carolina; Dan Ariely, associate professor at MIT's Sloan School of Management and chairman of Saffron's technical advisory board; and John Lynch, professor of marketing at Duke University's Fuqua School of Business.

One strategy suggested by the results, according to Ariely, might be to mix different techniques over time.“Initially providers can use collaborative-filtering approaches, but as they learn more about the individual user, they can use more and more of the individual approach,” he explains.

Alternatively, Web sites could use the individual-based approach to construct pseudo-users, or “bots,” and combine them with real customers in a collaborative-filtering system. That is the suggestion of John Riedl, an associate professor of computer science at the University of Minnesota and co-founder of Minneapolis-based Net Perceptions, a pioneer in commercializing collaborative filtering. Using a variety of bots to rate different items, he explains, provides a way to enrich sparse databases and quickly feed in information on new products.

In order to get the greatest benefit from the individual-based approach, however, companies must figure out how to identify and measure the product attributes that drive customers' preferences. That may be relatively straightforward for items such as laptop computers, but for categories such as music or books, it can be a formidable task.

Whatever the inner workings of the recommendation agent, it may improve performance to make the recommendation process more visible to the users. Providing insight into the reasoning behind the suggestions can increase the value and acceptance of recommendation agents, Riedl believes. Even better, if explaining the process makes it easier for customers to accept appropriate recommendations and reject inappropriate ones, it will also improve the quality of the feedback data that intelligent agents rely on.

But no matter how smart intelligent recommendation engines get, they will remain too complex for some companies' needs, observes Paul McNulty, vice president of marketing at Wheelhouse, a marketing-infrastructure-services company based in Burlington, Massachusetts. Companies with relatively few products, he points out, may find an older and often cheaper recommendation approach to be more efficient. Hand-coded business rules (business procedures that get programmed into the system in a fixed way, without the ability to adjust to learning) will remain an option.

If the problem of matching customers to products is relatively simple, then intelligent recommendation agents may be unnecessary, agrees Ariely. But the more complex systems can still pay off, he says, “when you as the seller don't have all the relevant knowledge and want to gain that knowledge, either from users about other users or from users about themselves.”

Reprint #:

42496

More Like This

Add a comment

You must to post a comment.

First time here? Sign up for a free account: Comment on articles and get access to many more articles.