A test of genetic algorithms in relevance feedback

Abstract

There have been recent applications of genetic algorithms to information retrieval, mostly with respect to relevance feedback. Nevertheless, they are yet to be evaluated in a way that allows them to be compared with each other and with other relevance feedback techniques. We here implement the different genetic algorithms that have been applied in the literature together with some of our own variations, and evaluate them using the residual collection method described by Salton in 1990 for the evaluation of relevance feedback techniques. We compare the results with those of the Ide dec-hi method, which is one of the traditional methods that yields the best results.

Genetic algorithms in relevance feedback: a second test and new contributions

“…As we mentioned above, in the present work we use a GA implemented for relevance feedback and optimized in earlier work (Lo opez-Pujalte, 2000) in which we experimented with the alternatives or improvements to the classical model taken from the literature and some of our own design (Lo opez- Pujalte et al, 2002a). In the following, we shall give some details of the characteristics of the GA that gave the best performance, and on which we shall implement each of the fitness functions defined in the previous section.…”

Structure of morphologically expanded queries: A genetic algorithm approach

“…IQBE [6] is a process in which the user does not provide a query, but document examples and the algorithms induce the key concepts in order to find other relevant documents. Mainly two information retrieval problems have been tackled with GAs: assigning weights to the query terms [21,28,23,14,13], and selecting query terms. Let us consider a number of proposals in the latter case, the one on which we focus our work.…”

“…However, this is not known at query time. Instead, it has been suggested in previous work to use the document scores as the fitness [14]. While this may not be intuitive, it turns out that variations of these scores after expansion are correlated with relevance [1].…”

Local search: A guide for the information retrieval practitioner

“…Sebastiani (2002) asserts that comparisons for classifiers can only be performed when the test collection is the same (identical documents, topics and relevance judgements), the same split between training and test set is used, and finally the same evaluation measure must be applied to the test set. Unfortunately this is not the case from the literature in local search in IR, for example Lopez-Pujalte et al (2002) only use 33 of the 225 Cranfield queries for their experiments. Drawing any conclusions from the literature on which is the best local search technique for any given task is therefore very difficult if not impossible (Sparck Jones and van Rijsbergen, 1976).…”

“…Others who have used this measure include Fan et al (2006) and Horng & Yeh (2000). Lopez-Pujalte et al (2002) and Lopez-Pujalte et al (2003b) provide an extensive survey on 12 fitness functions, and found that the choice of fitness function was essential when guiding GA's through the search space. Other examples include use of the Guttman model, a statistical measure of rank correlation which has many of the same properties of average precision according to Tamine et al (2003).…”