Web Page Streams and Relevance Propagation for Topic Distillation

Over the past decade, several studies in field of relevance propagation models have been proposed to improve quality of web search, which include hyperlink-based score propagation, hyperlink-based term propagation and popularity-based relevance propagation models; however, all of them have used low precision content similarity functions in the propagation process and their throughputs are not entirely satisfactory. In this paper, two stream-based content similarity functions that could be used to derive new relevance propagation models were introduced. In the proposed content similarity functions, the web page was split to different streams with different degrees of importance and the text of each web page was divided between these streams. To evaluate the proposed relevance propagation models, Letor 3.0 (including two standard web test collections) was used in the experiments. It was concluded that splitting web pages as different streams could provide significant improvement in relevance propagation models.