Emerging Applications of Link Analysis for Ranking

The booming growth of digitally available information has thoroughly increased the popularity of search engine technology over the past years. At the same time, upon interacting with this overwhelming quantity of data, people usually inspect only the very few most relevant items for their task. It is thus very important to utilize high quality ranking measures which efficiently identify these items under the various information retrieval activities we pursue. In this thesis we provide a twofold contribution to the Information Retrieval field. First, we identify those application areas in which a user oriented ranking is missing, though extremely necessary in order to facilitate a qualitative access to relevant resources. Second, for each of these areas we propose appropriate ranking algorithms which exploit their underlying social characteristics, either at the macroscopic, or at the microscopic level. We achieve this by utilizing link analysis techniques, which build on top of the graph based representation of relations between resources in order to rank them or simply to identify social patterns relative to the investigated data set. We start by arguing that Ranking Desktop Items is very effective in improving resource access within Personal Information Repositories. Thus, we propose to move link analysis methods down to the PC Desktop by exploiting usage analysis statistics, and show the resulted importance ordering to be highly beneficial for the particular scenario of Desktop Search. We then apply the same technique for Spam Detection. We connect people across email social networks based on their email exchanges and induce a reputation metric which nicely isolates malicious members of a community. Similarly, we model several higher level artificial constructs which could negatively manipulate generic link analysis ranking algorithms, and indicate how to remove them in the case of Web page ranking. Finally, we exploit manually created large scale information repositories in order to Personalize Web Search. We investigate two different types of such repositories, namely globally edited ones and individually edited ones. For the former category we project link analysis onto public taxonomies such as the Open Directory and define appropriate similarity measures which order the search output in accordance to each user’s preferences. For the latter one, we propose to expand Web queries by utilizing both text and link analysis on top of Personal Information Repositories. Extensive experiments analyzing both approaches show them to yield significant improvements over regular Google search.