Research Of Nearest Neighbor Classification Algorithm Based On Sample Selection

Nearest neighbor classification algorithms have been widely applied in various fields owing to its simple and effective features.Thus its disadvantages become much more obvious.Among this,complex data has led to huge time space consumption,which requires improved data quality by appropriate data preprocessing operations.And the demand for classification accuracy is constantly improving in practice,which requires constant updating of classification criteria and design of better classification algorithms.Therefore,this paper is research on the improvement of nearest neighbor classification algorithms.The main research contents are as follows.Firstly,we summarize several typical nearest neighbor classification and sample selection methods,analyze their respective algorithms and its advantages and disadvantages,and briefly describe the close relationship between sample selection and nearest neighbor classification.Then,the nearest neighbor classification methods take a lot of time and space,and its performance are seriously influenced by the noise samples.To overcome the issues,sample selection algorithm combined with mutual neighbors(MNSS)has been developed.The proposed MNSS uses the principle of the mutual neighbor to delete the noise samples and the redundant points in the original data set,so as to improve the quality of the data set and make the classifier more robust.Secondly,in order to improve the classification performance of nearest neighbor classification,the adaptive pseudo neighbor classification algorithm based on BP neural network(BPANN)that integrates the BP neural network and pseudo nearest neighbor classification rule(PNN)based on similarities between the within-class samples has been introduced.The mapping between the input and output of the BP neural network is used to adjust the distance weighting coefficient between the unlabeled sample point and its neighbors lying in each training set,thus making up this defect which the PNN can not obtain optimal distance-weighted value due to the distance weighting coefficient are determined subjectively.And the validity of the algorithm is verified by experiments.Finally,with respect to the classification results are interfered by the outliers,motivated by the ideas of PNN and the local mean principle,a pseudo nearest neighbor combined with local mean(LMPNN)has been developed.Through MNSS,LMPNN obtains the preprocessed data set,then utilizes the local mean information of multiple neighbors of test sample in each train set to avoid the influence of outliers on the classification performance.Comparative analysis through simulation and other methods,which shows the superiority of LMPNN algorithm.