The Study And Application Of Classification Algorithms For CRM

The rapid development of Data warehouse and data mining technologies have promoted the update of Decision Supporting System, also made a great change to the economic relationship between enterprises and customers. Customer Relationship Management (CRM), as a new application of data mining technologies in DSS recently, has made a breakthrough to the framework of traditional business management and merchandise planning. The traditional business strategy "all for production" is replaced by the brand-new concept "all for customer". So, people needs to identify, classify and evaluate the customers by understanding of the customer's characteristic and behaviors in CRM, and then fulfills the purposes of customer loyalty, customer profitability, and customer retention and customer acquisition through improving Customer Services. Aiming at the important part-customer classifier in CRM, the thesis adopts the decision tree classification algorithm to construct the customer classifier after analyzing the existing DM classifier tools. The thesis analyses and research the instability problem, efficiency problem and scalability problem which exist in many decision tree classification algorithms, then introduce a stable decision tree solution of using multiple attributes formed by almost best split and splitting predicate as the split point. Applied the solution to improved classification algorithms SLIQ and C4.5, the efficient, stable and scalable Classifier is achieved.As to SLIQ, the thesis uses almost best split point and splitting predicate instead of only one attribute's Gini as the split criterion. Thus, the stability of SLIQ algorithm has been improved. Furthermore, the generated decision tree would be more informative and concise. The improved SLIQ algorithm replaces the store-memory class list with database table, and makes a structural modification in some degree, which enabled to store more than one almost best split. Therefore, the improved SLIQ algorithm is free from the memory completely. The scalability is improved. As to C4.5, the thesis modifies the computing of the information gain in order to modularize the method to improve its instability. On the other hand, by replacing linear sorting with counting sorting to search the split point, the improved C4.5 algorithm is optimized in time complexity to some degree. Through improving algorithm in terms of common problems in decision tree algorithm, the thesis has achieved the purpose to construct the efficient, stable and scalable classifier. It also gives a way to improve the stability,efficiency, and scalability of the decision tree algorithm. The author constructed a CRM classifier with the improved algorithm, applying it in ChongQing Mobile Communication Company Phone-fee Management System. Building a classification module based on the customer attribute dataset, Empirical results illustrates that the trees constructed by the proposed algorithm are more stable,informative, and concise. The thesis proved the feasibility of improving the stability and scalability of the decision tree algorithms.