Jeff Bezos pinned great hopes on his company’s ability to create a “soul mate,” a mechanism that analyzes data generated by previous purchases and searches to suggest book and music titles that are likely to be interesting to visitors. In 2001, Bezos and others hoped that this technology, known as collaborative filtering, would increase the usefulness of Internet retailing and help
move e-tailers away from the discount bin and toward the value added sellers.
As computer processing power increased in the 1990s, researchers began to develop algorithms to help predict consumer behavior. The goal behind collaborative filtering was to replicate and automate the process of “word-of-mouth” recommendations by which people suggest products or services to one another. Collaborative filtering was particularly useful in helping users
choose between thousands or millions of options that were too complex to analyze individually.
Most collaborative filtering systems were comprised of a series of general steps. First, a large group of people’s preferences were registered. Using a similarity measure, a subgroup of people
were selected whose preferences were similar to the preferences of the person who sought advice. A weighted average of the preferences for that subgroup was then calculated. The resulting preference function was used to recommend options on which the advice-seeker had expressed no personal opinion as yet. If the similarity metric had indeed selected people with similar tastes, the chances were great that the options that were deemed desirable by that group would also be appreciated by the advice-seeker. An application typically recommended books, music CDs, or movies. More generally, the method could be used for the selection of
documents, services, or products of any kind.
The main liability with existing collaborative filtering systems was that it required the collection of preferences. In order to be reliable, most systems needed a large number of people (typically thousands) to express their preferences about a relatively large number of options (typically dozens).However, the system only became useful after a critical mass of opinions had been
collected. Users were not motivated to express detailed preferences in the beginning stages (e.g., by rating dozens of book titles on a 10 point scale), when the system could yet help them.
Amazon, for example, avoided this start-up problem by collecting preferences that were implicit in people’s actions. Customers who ordered books or music from Amazon implicitly expressed their preference for the titles they bought over the titles they did not buy. Customers who bought the same book or CD were likely to have similar preferences for other titles as well. Similarly, collaborative filtering could use customers’ behavior while surfing a site to infer their tastes. Search behavior, the amount of time a customer spent on a given product, and similar metrics could be used to indirectly infer a preference without requiring tedious data entry on the customer’s part.
Jeff Bezos pinned great hopes on his company’s ability to create a “soul mate,” a mechanism that analyzes data generated by previous purchases and searches to suggest book and music titles that are likely to be interesting to visitors. In 2001, Bezos and others hoped that this technology, known as collaborative filtering, would increase the usefulness of Internet retailing and help move e-tailers away from the discount bin and toward the value added sellers. As computer processing power increased in the 1990s, researchers began to develop algorithms to help predict consumer behavior. The goal behind collaborative filtering was to replicate and automate the process of “word-of-mouth” recommendations by which people suggest products or services to one another. Collaborative filtering was particularly useful in helping users choose between thousands or millions of options that were too complex to analyze individually. Most collaborative filtering systems were comprised of a series of general steps. First, a large group of people’s preferences were registered. Using a similarity measure, a subgroup of people were selected whose preferences were similar to the preferences of the person who sought advice. A weighted average of the preferences for that subgroup was then calculated. The resulting preference function was used to recommend options on which the advice-seeker had expressed no personal opinion as yet. If the similarity metric had indeed selected people with similar tastes, the chances were great that the options that were deemed desirable by that group would also be appreciated by the advice-seeker. An application typically recommended books, music CDs, or movies. More generally, the method could be used for the selection of documents, services, or products of any kind.The main liability with existing collaborative filtering systems was that it required the collection of preferences. In order to be reliable, most systems needed a large number of people (typically thousands) to express their preferences about a relatively large number of options (typically dozens).However, the system only became useful after a critical mass of opinions had been collected. Users were not motivated to express detailed preferences in the beginning stages (e.g., by rating dozens of book titles on a 10 point scale), when the system could yet help them.Amazon, for example, avoided this start-up problem by collecting preferences that were implicit in people’s actions. Customers who ordered books or music from Amazon implicitly expressed their preference for the titles they bought over the titles they did not buy. Customers who bought the same book or CD were likely to have similar preferences for other titles as well. Similarly, collaborative filtering could use customers’ behavior while surfing a site to infer their tastes. Search behavior, the amount of time a customer spent on a given product, and similar metrics could be used to indirectly infer a preference without requiring tedious data entry on the customer’s part.
การแปล กรุณารอสักครู่..
