Interview - ICT Projects in Serbia

Standard techniques for estimatingweb page relevance neglect the informationprovided by the visual layout of the page. Themost popular link-based technique, PageRank,uses a model of a random surfer to estimate theprobability that he visits a page at any giventime. This probability is assumed to beproportional to the relevance of the page.PageRank considers all outgoing links equallyprobable to be followed by the random surfer.Web designers use visual features to assigndifferent degrees of relevance to the links andhelp users in browsing and selecting the infor-mation they need. Discarding this informationmakes the random surfer assumption moreunrealistic and could have a strong negativeimpact in the accuracy of the predictedrelevance estimations.The proposed approach is based on thelayout analysis that assigns to each link in apage a set of visual features. This set offeatures describes the visual appearance of thelink and the logical context in which it appears.We created a model based on the real userobservations to represent how humans selectlinks based on their visual appearance. Finally,the user model and the visual information thatwe made available for each link are merged toestimate the probability that each single linkwill be followed in any given page. Theexperimental results showed that the visualPageRank improves the accuracy of therelevance estimation.