Research On Classification Methods And Its Application In Customer Recognition Of Banking

With the entry of foreign banks and the emergence of Internet finance,domestic banks are under enormous competitive pressure.In order to get the source of customers,the bank must first classify customers for identifying target customers,high-quality customers and potential customers,which means a lot to banks.Bayesian network algorithm is one of the most popular classification methods since it can learn to intuitive and interpretative structural model from data.Aimed at the characteristics of bank customer identification,thesis study the data preprocessing and process of building Bayesian models in detail,giving an effective process model for customer classification named ERNOK.Based on the real data test,thesis compares and analyzes the effect of different methods at the various stages of data discretizing,attribute selection and structure learning.Then the results of test and comparison verify the effect of ERNOK.Finally,based on the ERNOK method,the BN-based bank customer identification model is built,providing a reference for bank.Based on the real dataset for bank customer identification,the major tests are listed as follows:First,lots of experiments test on the large-scale data testify the effect on the classification performance of different methods at the various stages of data discretizing,attribute selection and structure learning,and thesis finds out the best approach at each stage.1.A large scale dataset is tested to compare the efficacy of Equal Width Discretization,Equal Frequency Discretization and Entropy Minimum Discretization.The statistical analysis of the experimental tests proves that the Entropy Minimum Discretization method works better than other methods significantly.2.Compare the efficacy of IG-based attribute selection method,GR-based attribute selection method and not using attribute selection method.Results of ANOVA analysis of variance and multiple comparison show that the attribute selection method based on information gain ratio performs best.3.Compare the effect of naive Bayes classifier,Tree Augmented Na?ve Bayes classifier and K2-based classifier.Statistical results show that in smaller data sets,the Naive Bayes classifier wins the best performance and in a larger scale data,K2 classifier wins the best performance.Second,considering of the data scales and combined with best practices for each stage,ERNOK,a bank customer recognition method based on Bayesian network,is proposed.The large-scale experiments test and comparison with other thesis' algorithm prove the validity of ERNOK.1.ERNOK preprocesses data by Entropy Minimum Discretization and GR-based attribute selection method.2.Different scale datasets uses different Bayesian structure learning methods.On a smaller scale datasets,thesis uses Naive Bayesian learning structure;on a larger data set,thesis uses K2 structure learning.3.Adding a threshold at forecast period effectively improve recognition rate of small class.Third,a Bayesian network based bank customer identification model is given and evaluated.1.In the model,factors such as work type,loan,mortgage,the credit default situation describe customers' basic information and other factors such as call duration,way of contact,number of contact,month,activity interval,the number of historical ties,last activity,relate to activities.The most important factors affecting customer classification are mortgage situation,month,call duration and the last activity result.2.A cost analysis to evaluate the effect of the application of ERNOK proves its validity.