Machine learning on web documents

The Web is a tremendous source of information: so tremendous that it becomes difficult for human beings to select meaningful information without support. We discuss tools that help people deal with web information, by, for example, blocking advertisements, recommending interesting news, and automatically sorting and compiling documents. We adapt and create machine learning algorithms for use with the Web's distinctive structures: large-scale, noisy, varied data with potentially rich, human-oriented features. Two standard classification algorithms, the slow but powerful support vector machine and the fast but inaccurate Naive Bayes, are modified to make them more effective for the Web. The support vector machine, which cannot currently handle the large amount of Web data potentially available, is sped up by “bundling” the classifier inputs to reduce the input size. The Naive Bayes classifier is improved through a series of three techniques aimed at fixing some of the severe, inaccurate assumptions Naive Bayes makes. Classification can also be improved by exploiting the Web's rich, human-oriented structure, including the visual layout of links on a page and the URL of document. These “tree-shaped features” are placed in a Bayesian mutation model and learning is accomplished with a fast, online learning algorithm for the model. These new methods are applied to a personalized news recommendation tool, “the Daily You.” The results of a 176 person user-study of news preferences indicate that the new Web-centric techniques out-perform classifiers that use traditional text algorithms and features. (Copies available exclusively from MIT Libraries, Rm. 14-0551, Cambridge, MA 02139-4307. Ph. 617-253-5668; Fax 617-253-1690.)