Previous work in hypertext classification has resulted in two principal approaches for incorporating information about the graph properties of the Web into the training of a classifier. The first approach uses the complete text of the neighboring pages, whereas the second approach uses only their class labels. In this paper, we argue that both approaches axe unsatisfactory: the first one brings in too much irrelevant information, while the second approach is too coarse by abstracting the entire page into a single class label. We argue that one needs to focus on relevant parts of predecessor pages, namely on the region in the neighborhood of the origin of an incoming link. To this end, we will investigate different ways for extracting such features, and compare several different techniques for using them in a text classifier.
展开▼
