Approach for constructing decision trees based on dispersion degrees

In the process of constructing a decision tree, the criteria of selecting partitional attributes will influence the classification accuracy of the tree. Therefore, the limitations of information entropy based approach and weighted mean roughness (WMR) approach are discussed, and a concept of conditonal attributes dispersion degrees in information systems is proposed. This concept is used to choose partitional attributes in the algorithm of decision trees construction. This approach can overcome the limitations of WMR approach. The results of experiments on the UCI datasets show that the precison retios of the decision trees constructed by using dispersion degree (DSD) are approximate to that of the ones constructed by using entropy-based approach. However, the time complexity of DSD approach is lower.