HITS AND LOGSOM ALGORITHMS

HITS AND LOGSOM ALGORITHMS
To date, index-based search engines for the Web have been the primary tools with which users searched for information. Experienced Web surfers can make effective use of such engines for tasks that can be solved by searching with tightly constrained keywords and phrases. These search engines are, however, unsuited for wide range of less precise tasks. How does one select a subset of documents with the most value from the millions that a search engine has prepared for us? To distill a large Wen-search topic to a size that makes sense to a human user, we need a means of identifying the topic's most authoritative Web pages. The notion of authority adds a crucial dimension to the concept of relevance: We wish to locate not only a set of relevant pages, but also those that are of the highest quality.

It is important that the Web consists not only of pages, but also hyperlinks that connect one page to another. This hyperlink structure contains an enormous amount of information that can help to automatically infer notions of authority. Specifically, the creation of a hyperlink by the author of a Web page represents an implicit endorsement of the page being pointed to. By mining the collective judgment contained in the set of such endorsements, we can gain a richer understanding of the relevance and quality of the Web's contents. It is necessary for this process to uncover two important types of pages: authorities, which provide the best source of information about a given topic and hubs, which provide a collection of links to authorities.

Hub pages appear in a variety of forms, ranging from professionally assembled resource lists on commercial sites to lists of recommended links on individual home pages. These pages need not themselves be prominent, and working with hyperlink information in hubs can cause much difficulty. Although many links represent the some kind of endorsement, some of the links are created for reasons that have nothing to do with conferring authority. Typical examples are navigation and paid advertisement hyperlinks. A hub's distinguishing feature is that they are potent conferrers of authority on a focused topic. We can define a good hub if it is a page that points to many good authorities. At the same time, a good authority page is a page pointed to by many good hubs. This mutually reinforcing relationship between hubs and authorities. At the same time, a good authority page is a page pointed to by many good hubs. This mutually reinforcing relationship between hubs and authorities serves as the central idea applied in the HITS algorithm (Hyperlink-Induced Topic Search) that searches for good hubs and authorities. The two main steps of the HITS algorithm are

Sampling component, which constructs a focused collection of web pages likely to be rich in relevant information, and

Weight-propagation component, which determines the estimates of hubs and authorities by an iterative procedure and obtains the subset of the most relevant and authoritative Web pages.

In the sampling phase, we view the Web as a directed graph of pages. The HITS algorithm starts by constructing the subgraph in which we will search for hubs and authorities. Our goal is a subgraph rich in relevant, authoritative pages. To construct such a subgraph, we first use query terms to collect a root set of pages from an index-based search engine. Since many of these pages are relevant to the search topic, we expect that at least some of them are authorities or that they have links to most of the prominent authorities. We therefore expand the root set into a base set by including all the pages that the root-set pages link to, up to a designated cutoff size. This set typically contains from 1000 to 5000 pages with corresponding links, and it is a final result of the first phase of HITS.

In the weight-propagation phase, we extract good hubs and authorities from the base set V by giving a concrete numeric interpretation to all of them. We associate a non-negative authority weight ap and a non-negative hub weight hp with each page p ∈ V. We are interested only in the relative values of these weights; therefore normalization is applied so that their total sum remains bounded. Since we do not impose any prior estimates, we set all a and h values to a uniform constant initially. The final weights are unaffected by this initialization.