Query-sensitive Feature Selection for Lazy Learners

Feature selection contributes to increasing many learners' accuracy by identifying and removing irrelevant features in multidimensional datasets. Conventional feature selection methods determine the optimal feature subset independently from and prior to the introduction of a new query. In general, some features will be relevant only in certain tasks. We argue that a query, as an indicator of the attention focus and current task, is a major part of the context and should be involved in the determination of the final feature subset. In this paper we attempt to propose a query-sensitive feature selection model, present two algorithms for applying such a feature selection method, and test their effectiveness by comparing their performances to those of the conventional selection algorithms. Our experiments are executed under a nearest neighbor classification environment and the results show a consistent improvement in the classification performance when a query-sensitive feature subset is selected and used for measuring similarities between the query and other instances. The results suggest that the performance of a lazy learner has the potential to increase through query-sensitive feature selection