Analyse de la robustesse des algorithmes de méta-recherche discriminante
This paper studies the sensitivity of four metasearch engines under different situations. The focus of this analysis is on trainable metasearch engines. Our main contribution is a large scale systematic analysis of the performance and behavior of these methods on several corpora. Firstly, we analyze how the choice and normalization of the relevance score delivered by base search engines influence the performance of metasearch. We then study the effectiveness of the metasearch engines on individual queries. We also analyze the robustness of the metasearch engines with regard to the variability in the training and test corpora. We finally present a preliminary analysis of active learning for query selection. All of these experiments demonstrate that the learned search engines are quite effective since they are uniformly better than both heuristic metasearch techniques and indivual engines. They are also particularly robust to changes in the training data sets and to the encoding of base search engines scores.