Collaborative Representation-Based Nearest Neighbor Classification

With the rapid development of information technology,a large number of useful or useless data are produced in all walks of life,so data mining and classification become increasingly important.In the face of explosive growth data,people need an effective method to analyze it.K-nearest neighbor(KNN)algorithm is the most classical method in the field of data mining and pattern recognition,which has been studied for decades.In this algorithm,the distance between samples is used to determine the nearest neighbor,and the classification result can be obtained only by adjusting the number of nearest neighbors k.therefore,KNN algorithm is simple and easy to implement.However,its classification performance is easily affected by noise points and k-value sensitivity.Collaborative representation(CR)classification is a new representation-based classification method.This algorithm uses all the training samples to express the test samples in linear collaboration,and has been proved to have good classification performance.However,when the different training samples are similar,the algorithm will lead to the effective training samples may not play a leading role in the classification process,resulting in classification errors.Based on the knowledge of collaborative representation and neighbor classification,this thesis proposes the following three research contents:(1)Collaborative Coefficient Based K-Nearest Neighbor Classifier(CCKNN)algorithm is proposed to improve the selective sensitives of nearest neighbors,especially in small sample size data sets with noisy points.CCKNN uses the collaborative representation coefficient corresponding to the training samples to select the nearest neighbor samples of the test sample points,and based on the collaborative representation coefficients corresponding to the neighbor points in each kind of neighborhood and the classification decision as the proposed method.Compared with the traditional distance method,the collaborative coefficients can reflect the similarity between samples better.Through the repeated experiments on real data sets and face data sets,the effectiveness of the proposed CCKNN algorithm has been verified.(2)Multi-Local Means Collaborative Representation-Based K-Nearest Centroid Neighbor Classifier(LMRKNCN)algorithm is proposed to reduce the probability of noise points that exist in near neighborhood and use the spatial distribution information of data,so as to improve the sensitivity of k-value selection in KNN algorithm.LMRKNCN first calculates the multi-local means of k nearest centroid neighbors,then uses the collaborative representation coefficients to represent the multi-local means linearly,and finally the classification decision is made based on the residuals of the test samples and the collaborative samples.The geometric distribution and local information around the test sample can be obtained by solving the local multiple means with the near centroid neighbor.On the other hand,the residual of the test sample and the cooperative representation can be used as the classification decision function,so that different local means near the centroid can make different contributions to the classification.The experimental results on real data sets and image data sets show the effectiveness of the proposed LMRKNCN algorithm.(3)Prototype system of neighbor image classification based on collaborative representation.Using Python as the development environment of the prototype system,the functions of the two algorithms mentioned in this article and the back-end modules are realized.The front-end interface adopts the HTML5 language,which makes the prototype system interface concise and has strong human-computer interaction.The classification results on the prototype system verify the practicability of the proposed algorithm.