CAJ | 학술논문

针对客户消费属性的多维、相关及不确定的特点,提出了基于粗糙神经网络（RS-NN）的客户消费分类模型。在揭示了客户消费分类问题的粗糙集特性基础上,设计出由预处理分类知识空间、建立消费分类模型、分类模型应用构成的研究框架,系统阐述了基于粗糙集的约简消费属性、提取分类规则、构建粗糙集神经网络初始拓扑结构、训练和检验网络模型等一系列关键技术,最后以某地区电信客户管理为建模示例。结果表明：RS-NN模型在模型结构、模型效率、分类预测精度方面均优于BP-NN算法,是一种有效和实用的客户分类新方法。
침대객호소비속성적다유、상관급불학정적특점,제출료기우조조신경망락（RS-NN）적객호소비분류모형。재게시료객호소비분류문제적조조집특성기출상,설계출유예처리분류지식공간、건립소비분류모형、분류모형응용구성적연구광가,계통천술료기우조조집적약간소비속성、제취분류규칙、구건조조집신경망락초시탁복결구、훈련화검험망락모형등일계렬관건기술,최후이모지구전신객호관리위건모시례。결과표명：RS-NN모형재모형결구、모형효솔、분류예측정도방면균우우BP-NN산법,시일충유효화실용적객호분류신방법。
The customer consumption classification topic is receiving increasing attention from researchers in the field of customer relationship management.The current research on customer consumption classification can be further improved in many areas.For instance,customer consumption classification models should take into consideration multidimensional and other related consumption attributes into classification analysis,avoidance of attribute redundancy,and selection of core classification attributes.Customer consumption models should identify input neurons,hidden layers and hidden neurons in order to reduce the complexity of classification structure and improve model＇s explanatory power.Existing classification methods are not effective at representing the inconsistency of consumption attributes and classes.JPThis paper proposed a customer consumption classification model by integrating rough set and neural networks based on the rough set-neural network（RS-NN） model.Rough set is the core theory underpinning this study.This paper reduced attribute values and adopted core consumption attributes in order to solve attribute redundancy and inconsistency problems.This paper also used customer classification rules and solved attribute inconsistency problems.In addition,by integrating classification rules into neural networks this paper constructed a classification and parameters to reduce the complexity of the existing consumption classification model and training time,and improve a user＇s learning,reasoning and classification abilities.This paper adopted Rosetta V1.4.41 and MATLAB to construct a customer consumption classification model.This proposed model includes customer knowledge reduction pre-process,classification network construction and classification application.In the pre-process of customer knowledge reduction,we conducted an unsupervised discretization method to process continuous consumption attributes.This method enabled us to process qualitative data in isometric conversion method,form consumption sheet and produce discernibility matrix of logical operations.We were also able to reduce attribute values,obtain core attributes,and extract reduced knowledge space via rough classification rules and credibility.In the consumption classification model,we assigned input neurons,output neurons,hidden layers and hidden neurons,identified the relationships between nodes and established the weight of attributes in order to construct the initial topology of the RS-NN model.We then normalized two data sets with training samples representing 80% of the population and 20% of test samples.This typology enabled us to run the model with BP algorithm.We adopted the Tansig function as transfer function,calculated the errors with generalized δ rule,and tested the precision of the model.In the consumption application system,we classified new customers with RS-NN model,and produced a prediction report for customer consumption behaviors.The paper shows that a RS-NN model is better than a BP-NN algorithm in customer structure,model efficiency and classification prediction accuracy.The RS-NN model is an effective and practical method for customer classification.