Clustering Algorithm Based On SOM And PAM And Its Application In Banking Customer Segmentation

Along with the development of Internet technology,all walks of life are undergoing tremendous changes,data analysis has penetrated into every corner of people's daily life.In the customer segmentation,how to classify the customer groups scientifically and accurately through various types of data has become an important issue for the banks and other industries.A good clustering algorithm can do segment quickly and accurately,it can provide strong support for marketing activities.In this paper,the improved clustering algorithm is applied to the bank customer segmentation.Considering the advantages and limitations of different clustering algorithms,and the effective of combination clustering algorithm,we present an improved clustering method(SOMPAM-SC)based on the combination of SOM(Self-Organizing Maps)and PAM(Partitioning Around Medoid)algorithms,and using the SC index to determine the optimal cluster number.Firstly,the artificial data sets(sample1?sample2)and UCI data sets(Liver Disorders?Breastcancer-wisconsin?Pima-indians-diabetes)are used as simulation objects to verify the validity of the SC index.Then UCI data sets(Iris?Waveform?Skin)are used to compare the accuracy of the proposed algorithm and other algorithms(PAM,SOM and SOM-Kmeans-SC).Finally,the results show that the SC index can accurately determine the optimal cluster number in different data sets,and the correctness of the algorithm is higher than other models.Synthesizes the classification accuracy,the running time and the validity of the clustering evaluation index,the effectiveness of PAM-SOM-SC is fully explained.Then,we apply the improved algorithm to the banking customer segmentation.Based on the data of the telephone marketing data of Portugal banking in UCI database,we select the customer's personal information,job information and other indicators to establish the SOM-PAM-SC model,and give some marketing suggestions.This not only solves the problem of large amount of data and noise data,but also avoids the blindness and subjectivity of the cluster number.