Research On Pseudo-Nearest Neighbor Classification Algorithm

Classification is one of the main tasks of data mining.Classification algorithms are to build a classification model through training data and to predict the samples without the decision tabels,which arc widely used in data analysis.The local men pseudo neighbor algorithm(LMPNN)combines the classical local mean K nearest neighbor algorithm(LMKNN)and the pseudo nearest neighbor algorithm(PNN),which makes full use of the information of each neighbor sample in a data set,and reduces the influence of noise points on classification accuracy.The main advantages of LMPNN are simple,fast and easy to implement.However,LMPNN also has some obvious defects.The main work of this paper is to overcome the defects and improve LMPNN,The specific research contents and results are as follows:Since the pseudo-nearest neighbor classification algorithm(LMPNN)is sensitive to the outliers and noise points in data,this paper trys to make the improvement on LMPNN and proposes a bidirectional selection-based pscudo oearest neighbor algorithm(BSPNN).Firstly,the k nearest neighbors are selected by using the proximity measure;and then,the test sample and the neighbor samples are selected bidirectionally through the mutual neighbor definition,Secondly,by calculating both the number of neighbors and the weighted distance of the local mean of the neighbors in each class,the Euclidean distance between the test sample and the pseudo-neighbor is obtained.Finally,the test sample is classified by the voting method which uses an improved class credibility measure.The proposed method has the advantages of being able to accuratly identify the noise points,reduce the sensitivity of the nearest neighbor number k,and improve the classification accuracy when dealing with complex classification tasks.The simulation experiments have been performed on 15 real data sets of UCI and KEEL,and KNN,WKNN,LMKNN.PNN,LMPNN,DNN as well as P-ICNN algorithms have been employed to compare with our proposed algorithm BSPNN.The experimental results show that the classification performance of BSPNN is significantly better than the most of the compard algorithms.A Parameter Independent Local Weighted Mean-based Pseudo Nearest Neighbor classification algorithm(PILWMPNN)is proposed to overcome the sensitively of the parameter k and ignores the different influence of each attribute on the classification results.Firstly,the latest variant of the differential evolution algorithm—Success-History based parameter Adaptation for Differential Evolution(SHADE),is used to optimize the training samples to obtain the best k value and a set of best weights related to the category.Secondly,the neighted distance between an testing sample and an sample in a catergory is computed by the obtained w,and the testing sample is then classified.Finally,the simulation experiments have been performed on 15 real data sets of UCI and KEEL,and KNN,FKNN,WKNN,LMPNN,LMKNN,MLMNN,WRKNN as well as WLMRKNN algorithms have been employed to compare with our proposed algorithm PILWMPNN.The classification accuracy and the value F1 increase about 28%and 23.1%,respectively.Furthermore,F-Measure,Wilcoxon signed ranks test,Friedman test and Holland-Wolfe test show that the proposed PILWMPNN outperforms the other eight algorithms in terms of classification accuracy and k value selection.