LAUSR

Instance selection aims to search for the best patterns in the training set and main instance selection methods include condensation methods, edition methods and hybrid methods. Hybrid methods combine advantages of both edition methods and condensation methods. Nevertheless, most of existing hybrid approaches heavily rely on parameters and are relatively time-consuming, resulting in the performance instability and application difficulty. Though several relatively fast and (or) parameter-free hybrid methods are proposed, they still have the difficulty in achieving both high accuracy and high reduction. In order to solve these problems, we present a new parameter-free hybrid instance selection algorithm based on local sets with natural neighbors (LSNaNIS). A new parameter-free definition for the local set is first proposed based on the fast search for natural neighbors. The new local set can fast and reasonably describe local characteristics of data. In LSNaNIS, we use the new local set to design an edition method (LSEdit) to remove harmful samples, a border method (LSBorder) to retain representative border samples and a core method (LSCore) to condense internal samples. Comparison experiments show that LSNaNIS is relatively fast and outperforms existing hybrid methods in improving the k -nearest neighbor in terms of both accuracy and reduction.