Research Of Feedback Clustering Algorithm

Cluster analysis is one of the basic methods in data analysis. It can be used as an independent data mining tools to obtain the distribution of datas, or as a pretreatment step for other data mining algorithms. Therefore,it has broad prospects in the domains of market or customer segmentation, pattern recognition, biology, space data analysis, Web and other documents classification ,etc. It has been widespread concerned by domestic and foreign scholars, and fruitful research results has achieved. However, conclusion from the practical application under current situation shows that most clustering algorithms consider only how to conduct a one-way analysis of raw data, and the effect is of little satisfaction.Based on this, first of all, a comprehensive analysis of the existing clustering algorithms and their application at home and abroad is made ,then their deficiencies are pointed out. Based on the traditional CURE algorithm, the idea of feedback clustering algorithm is put forward, which means feeding the result data of cluster analysis back to the the initial phase of cluster analysis, to adjust the process of cluster analysis based on feedback data.This process of result back to the cluster analysis stage is to be continued for refinement.In the feedback clustering algorithm, concepts of feedback sets, sets of relationship between two feedback sets and result sets are established innovatively,and the related character of these concepts are studied. Four steps of the feedback clustering algorithm are defined as result sets construction, result sets merging, cluster initializing, cluster post-treatment, and detail algorithm in each step is given. Using MyEclipse development tool, a simulation system is designed and implemented. Based on the simulation system, experiments are made from the aspects of algorithm complexity, clustering accuracy of the results ,and check rate of abnormal data for analysis and verification. The comparison with CURE algorithm is made,too.Finally, the feedback clustering algorithm is used in one of telecom operators's customer relationship management system (the total number of experimental data sets is 200,000) for customer segment.The results and analysis of the application shows that, the basic performance of the system is little affected, while it can continue finding data quality issues of the original CRM system, and in the cluster stage over disturb customers can be excluded out of the target customer. So the clustering accuracy of the results has a marked improvement.