Efficient Processing of Ranking Queries in Novel Applications

Ranking queries, which return only a subset of results matching a user query, have been studied extensively in the past decade due to their importance in a wide range of applications. In this thesis, we study ranking queries in novel environments and settings where they have not been considered so far. With the advancements in sensor technologies, these small devices are today present in all corners of human life. Millions of them are deployed in various places and are sending data on a continuous basis. These sensors which before mainly monitored environmental phenomena or production chains, have now found their way into our daily lives as well; health monitoring being a plausible example of how much we rely on continuous observation of measurements. As the Web technology evolves and facilitates data stream transmissions, sensors do not remain the sole producers of data in form of streams. The Web 2.0 has escalated the production of user-generated content which appear in form of annotated posts in a Weblog (blog), pictures and videos, or small textual snippets reflecting the current activity or status of users and can be regarded as natural items of a temporal stream. A major part of this thesis is devoted to developing novel methods which assist in keeping track of this ever increasing flow of information with continuous monitoring of ranking queries over them, particularly when traditional approaches fail to meet the newly raised requirements. We consider the ranking problem when the information flow is not synchronized among its sources. This is a recurring situation, since sensors are run by different organizations, measure moving entities, or are simply represented by users which are inherently not synchronizable. Our methods are in particular designed for handling unsynchronized streams, calculating an object's score based on both its currently observed contribution to the registered queries as well as the contribution it might have in future. While this uncertainty in score calculation causes linear growth in the space necessary for providing exact results, we are able to define criteria which allows for evicting unpromising objects as early as possible. We also leverage statistical properties that reflect the correlation between multiple streams to predict the future to provide better bounds for the best possible contribution of an object, consequently limiting the necessary storage dramatically. To achieve this, we make use of small statistical synopses that are periodically refreshed during runtime. Furthermore, we consider user generated queries in the context of Web 2.0 applications which aim at filtering data streams in forms of textual documents, based on personal interests. In this case, the dimensionality of the data, the large cardinality of the subscribed queries, as well as the desire for consuming recent information, raise new challenges. We develop new approaches which efficiently filter the information and provide real-time updates to the user subscribed queries. Our methods rely on a novel ordering of user queries in traditional inverted lists which allows the system to effectively prune those queries for which a new piece of information is of no interest. Finally, we investigate high quality search in user generated content in Web 2.0 applications in form of images or videos. These resources are inherently dispersed all over the globe, therefore can be best managed in a purely distributed peer-to-peer network which eliminates single points of failure. Search in such a huge repository of high dimensional data involves evaluating ranking queries in form of nearest neighbor queries. Therefore, we study ranking queries in high dimensional spaces, where the index of the objects is maintained in a purely distributed fashion. Our solution meets the two major requirements of a viable solution in distributing the index and evaluating ranking queries: the underlying peer-to-peer network remains load balanced, and efficient query evaluation is feasible as similar objects are assigned to nearby peers.