The Research On Decision Tree Algorithm Based On Rough Set And Application In CRM

Data mining is a kind of technology that can extract the valuable information from the massive data. It can be used to find the cluster pattern, the association pattern, and the classification pattern, hidden in the data, in order to help people decision-making. The classification is an important task of data mining, and the decision tree is a method used effectively for data classification. The structure of the decision tree is simple. The decision tree is easy to produce rules and to be understood. It is used widely in practical applications.Firstly, this paper improves the decision tree algorithm based on rough set theory. Based on rough set theory, it reduces attributes of the decision-making table and then constructs the decision tree. The capacity of handling the noise data based on classic rough set is insufficient. Therefore, based on the variable precision rough set thinking, this paper proposes the decision tree algorithm based on the variable weighted average roughness, which improve the algorithms presented by Dr.Jiang Yun et al (2004).The experimental results show that the decision tree based on the improved algorithm is simpler in the structure and its extensive ability is stronger.Secondly, this paper will present decision tree algorithm based on the discern degree of attributes. The main idea is during every branch of the decision tree the data belong to different categories Separate as much as possible and the data belong to the same categories pool together as much as possible. Similarly, Simulation results show that in most data sets this decision tree algorithm is better than the classic ID3 algorithm.Finally, with the customer relationship management (CRM) in telecommunications industry as the background, this paper builds the churn of early warning models, based on the basic data mining process. These models are built based on the ID3 algorithm and the algorithms that we improve and present in a chum of Telecom data set. The result shows that the models constructed with the latter two algorithms are better than the former.