Distributed Information Retrieval For Disruption-Tolerant Mobile Networks

We design and evaluate a distributed information retrieval system that operates over a mobile network where a wire- less infrastructure unavailable. Such networks are common in developing nations, disaster-stricken areas, and even in the rural areas of the technologically progressive countries. This poses a new challenge for distributed IR, which nor- mally relies on a wired Internet or always-available wireless coverage among mobile peers. In our mobile system, queries are propagated among peers only as they intermittently are in wireless range of one another. For each query received, peers retrieve top-ranked documents from their local collec- tion and send them to the source of the query. Intermediate peers on the path to the source have to manage a finite buer filled with documents from multiple collections and multiple queries. When too many documents are in the system, the intermediate peers must drop documents that are either un- likely to be relevant or for which a successful path to the destination is unlikely. To enable such a system, we propose a score normalization technique that works across queries and across multiple collections. We show that our method returns more relevant documents in the mobile network than existing normalization methods, which are not intended for multiple queries. Additionally, we compare our approach to existing networking algorithms for delivering data in such challenged networks. We show that although our method delivers less total documents, it delivers significantly more documents that are relevant to sources of queries.