A New Searching Method of Splitting Threshold Values for Continuous Attribute Decision Tree Problems

In the paper, we extend the well-known golden-section search (GSS) method to make an unprecedented attempt to do discrete sequence searches. The GSS method is originally used to find the extremum of a strictly unimodal continuous function. We apply it on searching the best threshold for discretizing continuous attribute data in decision tree problems. Compared to typical methods, the shortcomings relating to massive calculation requirements for searching threshold values are eliminated. Whether it is used along with information gain or Gini index as the measure indicator for data purity of decision tree, the algorithm produces good results. To verify the proposed method, data set provided by UCI database is used on Mat lab platform to carry out the simulation. Results indicate that under the same performance index, the discrete GSS method significantly lowers iteration numbers of searching threshold values and, hence, verify the feasibility of this algorithm.