Research On Clustering Algorithm Based On Boundary Recognition

Most of the existing clustering algorithms need to define the centers or find the main bodies of the clusters.These algorithms may have some common limitations:(1)The number of clusters needs to be given by users.(2)The heterogeneity of clusters cannot be processed properly.(3)Parameters in the algorithms are difficult to determine.This paper proposes clustering algorithms based on boundary recognition,which can better overcome the above limitations:(1)After determining the boundary between clusters,they can be extracted one by one,so the number of clusters is no need.(2)Using datadriven methods to determine the boundary between clusters,can accurately find the boundary between heterogeneous classes.(3)The boundary between clusters is determined by the distribution of data.The methods proposed in this paper have no parameters,or there are few parameters.In the clustering problem,it is easy to determine the center and the main bodies of the clusters,but it is difficult to find the boundary of the clusters,which also requires a lot of calculation.Therefore,according to the manifestations of the boundary between clusters under different theoretical frameworks,this paper applies different processing methods to the boundary to achieve efficient boundary recognition,so as to complete the task of clustering.The first work of this paper is to propose a density-based boundary projection clustering algorithm.Aiming at the simplest spherical clustering problem,this algorithm first uses the fast mean shift algorithm to determine the center of a cluster,then uses the boundary projection technique to project the distribution of sample points in the multidimensional space into the one-dimensional distance space to determine the radius of the cluster,and finally extracts the clusters one by one according to the center and radius of clusters.The second main work of this paper extends the density-based boundary projection clustering algorithm to non-spherical clustering problems and proposes a pathbased boundary transformation clustering algorithm.In this algorithm,path distance is used instead of Euclidean distance to transform irregular boundaries between non-spherical clusters into regular spherical boundaries,and then the frame of density-based boundary projection clustering algorithm can be used to complete the clustering task.Although the above two methods have the ability to anti-noise for a certain degree,boundary needs accurate refinement when the condition is much more complex.The third main work of this paper further improves the efficiency of boundary recognition and proposes a gridbased boundary sharpening clustering algorithm.This algorithm uses the grid to divide the space where the sample points are located into cells,and then adaptively determines the threshold value according to the number of sample points distributed in the cells.The threshold value is used to divide the cells into the foreground grid representing the main body of the class and the background grid representing the boundary between the classes,and then uses the mathematical morphology method to sharpen the boundary,so as to obtain the accurate boundary between the clusters.The validity,robustness and expansibility of the proposed algorithm are verified by sufficient reference data sets.In addition,real data sets from different fields are used to verify the universality of the algorithm.Experimental results show that the proposed clustering algorithm based on boundary recognition can efficiently perform clustering tasks on data sets of different distribution types.