Abstract

Ranking search results is essential for information retrieval and Web search. Search engines need to not only return highly relevant results, but also be fast to satisfy users. As a result, not all available features can be used for ranking, and in fact only a small percentage of these features can be used. Thus, it is crucial to have a feature selection mechanism that can find a subset of features that both meets latency requirements and achieves high relevance. In this paper we describe a 0/1 knapsack procedure for automatically selecting features to use within Generalization model for Document Ranking. We propose an approach for Relevance Feedback using Expectation Maximization method and evaluate the algorithm on the TREC Collection for describing classes of feedback textual information retrieval features. Experimental results, evaluated on standard TREC-9 part of the OHSUMED collections, show that our feature selection algorithm produces models that are either significantly more effective than, or equally effective as, models such as Markov Random Field model, Correlation Co-efficient and Count Difference method.