Feature selection for query classification & candidate set selection for selective information retrieval based on nature inspired evolutionary algorithms

When studies in the selective information retrieval systems literature are examined, it is seen that predefined candidate sets are used in all of them. However, there is no clear information on what kind of methods were used to determine these candidate groups. In this thesis study, the process of determining the candidate set for selective information retrieval systems and selecting the attributes for query classification was carried out by using evolutionary algorithms within the scope of joint optimization. Experimental studies are evaluated in three groups: Selective Stemming, Selective Term Weighting, and Selective Stemming and Term Weighting, where both approaches are considered together. In the first two approaches, the selection is made among alternative models of a component, while all other components are kept fixed. In the third one, all subsets obtained by pairwise matching of models belonging to different components are considered as separate information retrieval systems. The k-fold cross validation method was used to verify experimental studies. It has been observed that the proposed method achieves higher performance than the single system, which is the best average for each data set. Additionally, the results found were statistically significant. The components and features encoded in the genotypes of the individuals ultimately obtained were examined within the scope of information retrieval perspective.