The Method Of Instance Selection In Pattern Classification

Data selection plays a great role in pattern recognition. In studying system, we note that boundary data and redundant data which have seriously reduced the recognition rate of the sample become our problems to be solved. This article focuses on the instance selection. For different problems, we improve two original methods, and give one new view, and the author's main production is contained as following:(1). Based on the feature attribute, we propose a novel recognition method to avoid the adverse influence of noisy data in the training set. The main idea of this method is to find k nearest neighbors in each feature instead of k nearest neighbors in all samples.(2). The edited technique usually involves two sets:the testing set and reference set. We Exchanged the editing set and testing set, and improved the original method.(3). Based on removing redundant data and retaining reliable data, we propose a new method to complete instance selection. The main idea of this method is that we compare classes in initial training set with clusters after K-means clustering. We make the intersection between every class and cluster as the reliable samples. In addition, to indicate different samples'different importance in each class, we give a fuzzy relationship in the reliable data.In the end, the author summarizes the research process and points out some limitations of this paper.