analyze the semantic content of Web pages. However, the ability to use this eigenvector approach with real data that are random or contain much uncertainty is the key to PageRank. Within a few years, everybody was using Google, and “to google” had become a verb. When Brin and Page’s company went public in 2004, its initial public stock offering raised $27 billion.

In many other applications, finding eigenvectors through SVD has proved to be effective for aggregating the collective wisdom of humans. From 2006 to 2009, another hot Internet company ran a competition that led to a number of advances in this field.

image

Within a few years, everybody was using Google, and “to google” had become a verb.

Netflix, a company that rents videos and streams media over the Internet, had developed a proprietary algorithm called Cinematch, which could predict the number of stars (out of five) a user would give a movie, based on the user’s past ratings and the ratings of other users. However, its predictions were typically off by about 0.95 stars. Netflix wanted a better way to predict its customers’ tastes, so in 2006 it offered a million-dollar prize for the first person or team who could develop an algorithm that would be 10 percent better (i.e., its average error would be less than about 0.85 stars). The company publicly released an anonymized database of 100 million past ratings by nearly half a million users so that competitors could test their algorithms on real data.

Rather unexpectedly, the most effective single method in the competition turned out to be good old-fashioned SVD. The idea is roughly as follows: Each customer has a specific set of features that they like in a movie—for instance, whether it is a drama or a comedy, whether it is a “chick flick” or a “guy flick,” or who the lead actors are. A singular value decomposition of the database of past ratings can identify the features that matter most to Netflix customers. Just as in the genomics example, the mathematical sciences cannot say what the features are, but they can tell when two movies have the same constellation of factors. By combining a movie’s scores for each feature with the weight that a customer assigns to those features, it can predict the rating the customer will give to the movie.

The team that won the Netflix Prize combined SVD with other methods to reach an improvement of just over 10 percent. Not only that, the competition showed that computer recommendations were better than the judgment of any human critic. In other words, the computer can predict how much your best friend will like a movie better than you can.

The above examples attest to the remarkable ability of eigenvector methods (often in combination with other techniques) to extract information from vast amounts of noisy data. Nevertheless, plenty of work remains to be done. One area of opportunity



The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement