Study On Cluster Analysis And Outlier Detection Based On Natural Neighbor And Local Resultant Force

With the development of Internet information technology,the collection of large-scale data is more and more convenient,and the structure of data is more and more complex.It is very difficult to label a large number of complex datasets.How to mine valuable information from complex unabled data has become the research focus of unsupervised learning.Clustering analysis and outlier detection are two very important research directions of unsupervised learning and have been widely used,such as image segmentation,face recognition,credit fraud detection,network instrusion detection and other fields.The concept of density core has been put forward,which makes the clustering algorithm based on density core performs well in dealing with spherical,complex manifold and datasets of varying densities.DCore algorithm is one of the most representative clustering algorithms based on density core.However,there are two defects in DCore algorithm,one is that the parameters are difficult to set and the other is that the algorithm is still unable to cluster data with extremely large variations in density.In order to overcome the defects of DCore algorithm,we propose a novel density core-based clustering algorithm with the local resultant force(DCLRF).Firstly,we design a centrality value(CE)based on natural neighbor theory and local resultant force to identify core points and non core points,and we use CE to extract core points.Then,we use natural neighbor structure of core points to get the clustering results.Experiments show DCLRF algorithm can deal with datasets containing spherical,complex manifold and extremely large variations in density without setting any parameters manually.In response to the problem that LOF algorithm can't detect outliers in datasets containing linear,complex manifold or large variations in density.We propose a novel outlier detection algorithm based on neighborhood weighted(NWOD).Firstly,we propeses the concept of weighted local density based on natural neighbor theory and weighted neighborhood graph.Then,the corresponding outlier scores are obtanied by comparing the weighted local density difference between the data objects and it's neighbors.Finally,the outlier scores are sorted from large to small.The bigger the value of outlier scores is,the greater the possibility of data object is outlier.Experiments show NWOD algorithm can detect outliers in datasets containing linear,complex manifold or extremely large variations in density without setting neighbor number manually.