Hits on the web: how does it compare?

ABSTRACTThis paper describes a large-scale evaluation of theeffectiveness of HITS in comparison with other link-based rankingalgorithms, when used in combination with a state-of-the-art textretrieval algorithm exploiting anchor text. We quantified theireffectiveness using three common performance measures: the meanreciprocal rank, the mean average precision, and the normalizeddiscounted cumulative gain measurements. The evaluation is based ontwo large data sets: a breadth-first search crawl of 463 millionweb pages containing 17.6 billion hyperlinks and referencing 2.9billion distinct URLs; and a set of 28,043 queries sampled from aquery log, each query having on average 2,383 results, about 17 ofwhich were labeled by judges. We found that HITS outperformsPageRank, but is about as effective as web-page in-degree. The sameholds true when any of the link-based features are combined withthe text retrieval algorithm. Finally, we studied the relationshipbetween query specificity and the effectiveness of selectedfeatures, and found that link-based features perform better forgeneral queries, whereas BM25F performs better for specificqueries.