Data reduction has been an important issue of data mining including data compression

data adjustment

feature extraction

and so on

however

existing methods of data reduction mainly focus on reduction of features and dimensions

methods of reduction to the number of samples always limit to specific data sets which lack of generality. Aiming at general feature of data distribution in data sets

define a new kind of measurement based on opening angle. This measurement can distinguish essential difference of distribution between kernel objects and boundary objects

and realize data compression which takes kernel objects as center for data sets. By data reduction and test on 20 typical simple sets which have different features on UCI public test platform

the result demonstrates the proposed method can extract kernel objects in data sets effectively; by using the typical kmeans algorithm to cluster the data sets before data reduction

cluster accuracy of reduced data sets is apparently higher than that of original data sets.