Please use this identifier to cite or link to this item:
|Title:||Relevance prediction in similarity-search systems using extreme value theory|
Torres, Ricardo da Silva
|Abstract:||Among the challenges present in the design of retrieval systems, how to accurately assess their performance is perhaps one of the most important. Many applications such as rank aggregation or relevance feedback can be significantly improved with online effectiveness estimation of queries. Thus, developing methodologies that can estimate performance with minimal supervision and at query time is of utmost importance for improving the results of existing retrieval systems. In this work, we explore score-based, post-retrieval approaches for relevance prediction of search systems. We first introduce two statistical methods based on the Extreme Value Theory to estimate which of the objects retrieved for a query are relevant. Our prediction approach uses this estimation as a method to infer the overall performance of a query. The two relevance prediction methods were evaluated in image datasets covering several modalities and scoring approaches. We conducted experiments comparing the ground-truth relevances of several ranks with predictions generated by our proposed approach, measuring their effectiveness by way of normalized accuracy and Matthews Correlation Coefficient. Furthermore, we also evaluate the precision deducted from our approaches with the system’s expected performance. Those experiments show that the proposed approaches succeed in most relevance prediction scenarios of the top-ranked objects of a query, obtaining high accuracy|
|Subject:||Distribuição de Weibull|
Organização da informação
|Appears in Collections:||IC - Artigos e Outros Documentos|
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.