Research Of The Density-based Clustering Algorithm

As a basic method in unsupervised learning,clustering has been widely used in various fields,such as artificial intelligence,data analysis,image processing,biomedicine and so on.At present,many clustering algorithms have been proposed according to different clustering ideas.However,these algorithms suffer from several inevitable drawbacks.Firstly,most clustering methods require the cluster number in advance.Secondly,some methods show the serious sensitivity on the clustering parameters.Thirdly,some methods cannot well solve the high dimensional datasets and produce poor clustering results.Moreover,there are some methods which cannot accurately discover the clusters with different shapes and different densities.In order to overcome these defects,we make a study on the clustering algorithm based on analysis of sample density.The innovative achievements are summarized as follows:(1)In order to improve the robustness and effectiveness for noise data,a clustering algorithm based on density difference is proposed to realize automatic clustering.Firstly,according to different densities of noise data and useful data,the proposed algorithm realizes noise removal and classification of useful data based on density difference.By constructing neighborhoods between data,the classification of different classes among useful data is further realized.Finally,the effectiveness of the proposed algorithm is demonstrated by simulation experiments.(2)For clustering the useful data with different densities,an automatic clustering algorithm based on region segmentation is proposed.Firstly,we introduce a new density based on reverse nearest neighbor,which can detect dense points and sparse points in sub-regions effectively and divide the data set into several sub-regions.Then,the merging criterion can automatically merge the relevant regions without the number of clustering and threshold limitation.Finally,the experimental results fully demonstrate the effectiveness and feasibility of the proposed algorithm.(3)In order to solve the problem of high-dimensional data set clustering,a deep clustering algorithm combining neural network and density ordered graph is proposed.Firstly,a pre-training self-coding network is constructed,and then a density ordered graph is obtained based on coding features.The loss function of the defined algorithm includes two parts:the reconstruction loss of the self-coding network and the loss of the density ordered graph.The gradient descent algorithm is used to optimize the network parameters until the terminate condition is reached.Finally,the experimental results demonstrate that the proposed algorithm is effective and feasible.