Selective web information retrieval

One of the main challenges in Web information retrieval is the number of different retrieval approaches that can be used for ranking Web documents. In addition to the textual content of Web documents, evidence from the structure of Web documents, or the analysis of the hyperlink structure of the Web, can be used to enhance the retrieval effectiveness. However, not all the queries benefit equally from applying the same retrieval approach. An additional challenge is posed by the fact that the Web enables users to seek information by searching and browsing. Therefore, users do not only perform typical informational search tasks, but also navigational search tasks, where the aim is to locate a particular Web document, which has already been visited before, or which is expected to exist. In order to alleviate these challenges, this thesis proposes selective Web information retrieval, a framework formulated in terms of statistical decision theory, with the aim to apply an appropriate retrieval approach on a per-query basis. The main component of the framework is a decision mechanism that selects an appropriate retrieval approach on a per-query basis. The selection of a particular retrieval approach is based on the outcome of an experiment, which is performed before the final ranking of the retrieved documents. The experiment is a process that extracts features from a sample of the set of retrieved documents. This thesis investigates three broad types of experiments. The first one counts the occurrences of query terms in the retrieved documents, indicating the extent to which the query topic is covered in the document collection. The second type of experiments considers information from the distribution of retrieved documents in larger aggregates of related Web documents, such as whole Web sites, or directories within Web sites. The third type of experiments estimates the usefulness of the hyperlink structure among a sample of the set of retrieved Web documents. The proposed experiments are evaluated in the context of both informational and navigational search tasks with an optimal Bayesian decision mechanism, where it is assumed that relevance information exists. This thesis further investigates the implications of applying selective Web information retrieval in an operational setting, where the tuning of a decision mechanism is based on limited existing relevance information and the information retrieval system's input is a stream of queries related to mixed informational and navigational search tasks. First, the experiments are evaluated using different training and testing query sets, as well as a mixture of different types of queries. Second, query sampling is introduced, in order to approximate the queries that a retrieval system receives, and to tune an ad-hoc decision mechanism with a broad set of automatically