Study On Bank Customer Segmentation Based On Improved SOM

With the rapid development of economy and the continuous progress of science and technology,the competition in the banking industry is becoming increasingly fierce.Its competitive point has gradually changed from "product-centered" to "customer-centered".In order to better meet the needs of users and provide better user experience,it has become the key to the development of banks.At this time,customer segmentation has become particularly important.However,bank customer data is complex and huge.How to build a customer segmentation model by clustering algorithm,effectively and accurately divide a large number of bank customers,dig out the potential needs of customers,and sell the right products to the right customers at the right time has become the focus of banking research.Because of the characteristics of high-dimensional bank customer data,huge amount of data and noisy data,this paper proposes an improved clustering algorithm SOM-Kmedoids-CH for customer segmentation.The main research contents are as follows:Because of the characteristics of high-dimensional bank customer data,huge amount of data and noisy data,this paper proposes an improved clustering algorithm SOM-Kmedoids-CH for customer segmentation.The main research contents are as follows:(1)Self-organizing neural network(SOM)is effective in processing large sample data,but the selection of learning rate greatly affects the clustering effect of SOM neural network.When the learning rate is large,the weight vector will oscillate and update repeatedly,resulting in the decline of learning stability.When the learning rate gradually approaches zero,although the learning stability of the network is improved,the convergence speed of the network will decrease.To solve this problem,this paper sets the learning rate as a monotonically decreasing function with respect to time t,which can ensure that the model can be learned at a faster speed at the beginning of training,and at the end of training,the learning rate approaches zero,thus ensuring the stability of model training.Thereby improving the clustering efficiency of the algorithm.(2)Considering that the future development trend of subdivision technology is to combine the advantages of different algorithms,this paper combines the improved SOM neural network with the fast K-center algorithm,and uses CH index to judge the best cluster number.Firstly,the improved SOM neural network is used to train sample data to output prototype vector,which is much smaller than the original sample data and keeps the original topological structure unchanged.The prototype vector output by SOM is used as the input of fast K center point for clustering.At the same time,the CH index is used to evaluate each clustering result,and the K value corresponding to the highest CH value is selected as the best clustering center.This method solves the problems of the traditional SOM algorithm,such as long convergence time,blindness of artificially determining the number of clusters at the fast K center point and insensitivity to noise data.In this paper,UCI data set is used to test the convergence speed of the improved SOM,then the validity of CH index is verified,and then the algorithm in this paper and other algorithms are tested and compared.From the experimental results,the convergence rate of the improved SOM is faster than that of the traditional SOM,and the CH index can accurately judge the best cluster number.The algorithm in this paper is superior to other clustering algorithms from three aspects: running time,classification accuracy and whether the best clustering number is effective.Then,this paper takes the customer data of Kaggle Bank in Santander,official website as a sample,uses the improved SOM-Kmedoids-CH clustering algorithm to model,divides the bank customers into four categories,makes cluster analysis on the customers,and provides corresponding opinions and suggestions for different types of bank customers.