Study On Incremental Attribute Clustering Method Based On Multi-granularity

Cluster analysis can obtain valid information by dividing data,which has been pervasively used in transportation,industry and other related fields.The traditional twoway clustering representation cannot clearly represent the objects belong to the fridge region of the clusters.Three-way clustering representation can reveal the three relationships between an object and a cluster by using a pair of sets as the core region and the fringe region.However,most of the current clustering algorithms can only deal with static datasets.Furthermore,there is little research on attribute incremental data clustering.Therefore,this thesis carries out the following research work based on granular computing,three-way clustering and other methods.Aiming at solving the clustering problem of attribute incremental data,this thesis proposes a multi-granularity incremental attributes clustering method.Firstly,the algorithm obtains the initial clustering result by the density peak algorithm.Secondly,a multigranularity layer was formed for a new set of attribute particles at a certain moment by combined with the original attribute particles.Finally,on the premise of not repeating the clustering,the clustering result is dynamically updated by combined the neighborhood information and the original clustering result of the object until there is no new attribute set is added.Aiming at the clustering of attribute incremental data with uncertain information,the previous method is improved and a three-way clustering method based on multi-granularity incremental attributes is proposed.The algorithm consists of five steps.Firstly,the algorithm obtains the initial three-way clustering results by the improved density peak algorithm.Then,for a new set of attribute granules at a certain time,the redundant attributes are filtered firstly and the remaining attributes are adding into the original set of attributes.And,the mean distances between fringe points and other non-fringe points are judged.After that,neighborhood distance information and neighborhood attribution cluster information are counted and calculated,in which neighborhood objects belong to the core/fringe region of the clusters.And,combining the original clustering results and Markov distance to dynamically update the original clustering ascription.Then,the cluster ascription is divided into core region and fringe region based on the idea of three-way decisions,and then the change of number of the clusters is judged.Finally,the algorithm is end until no new set of attribute granules is added.In this thesis,10 UCI real data sets,such as Iris,Statlog and Waveform,are used to validate the effectiveness of the proposed algorithm.The experimental results show that the proposed method is superior to the contrast method in most cases on the indices of NMI,RI and Accuracy,and indicate that the research work in this thesis is effective for dealing with the problem of attribute incremental data.