Research On Fuzzy Semanteme Of Decision Trees Algorithms

Although decision trees algorithms based on semanteme can make data and concepts match more precisely, they have rigid division defects. In the process of getting semanteme, the values of continuous attributes in the training set are mapped to certain concepts in semantic concept classes, not considering their memberships of certain concepts. However, whether a data belongs to certain concept is ambiguous.This is the fuzzy problem of semanteme. Fuzzy logic solves the ambiguity problem to some extent, but its essence is still using specific membership to represent uncertain fuzziness, so fuzzy logic is not thorough.To solve the problem of decision trees algorithms based on semanteme, this article researches on fuzzy semanteme of decision trees algorithms, and proposes a method for random fuzzy semanteme of continuous attributes named CARFS. By utilizing semantic concept trees and fuzzy c-means algorithm to get memberships of continuous attributes values, and taking advantage of cloud model to obtain accuracies of memberships simultaneously, that is randomness, we solve problems that fuzziness is not taken into account in decision trees algorithms based on semanteme and fuzziness is not thorough. This article proposes a method generating semantic concept trees automatically. By improving discretization of real-value attributes based on genetic algorithm, form leaf nodes of semantic concept trees, and then apply unweighted pair-group method with arithmetic means (UPGMA) to promote hierarchies, which can make full use of each value of continuous attributes and make hierarchies of concept trees much more reasonable. After that, we introduce CARFS method to Fuzzy ID3 algorithm, and propose a decision trees algorithm based on fuzzy semanteme named SFID3 and point out how to find the optimal decision tree.We select Iris, Adult and Wine datasets provided by UCI as the training sets. The experiment results prove that the SFID3 algorithm is feasible and effective. The experiments prove that decision trees constructed by SFID3 have higher correct classification rate compared with decision trees algorithms based on semanteme.