Models for Metasearch

Abstract

Given the ranked lists of documents returned by multiple search engines in response to a given query, the problem of metasearch is to combine these lists in a way which maximizes the quality of the combination. In this talk, we describe two different models for the metasearch problem: one based on Decision Theory resulting in a Bayesian model and another based on Social Choice Theory resulting in a multi-candidate election model. We describe efficient algorithmic implementations of metasearch strategies in each of these models, and our experimental results demonstrate that these algorithms generally outperform existing ad hoc techniques on benchmark data sets. Finally, we describe a constrained oracle model for deriving upper bounds on the potential performance of metasearch algorithms, and we compare and contrast the performance of our techniques with the limits on metasearch performance thus derived.