Improved Decision Tree Based on Double-attribute and Part-match
HE Tian-zhon · 2015
Decision tree algorithm only chooses one attribute as the root or the node attributes, even if there are two best attributes. So the decision tree algorithm produces few rules. Using full-match method, a new instance only matches one rule or none, which decrease the accuracy of the decision tree. Aiming at the problem, we improve decision tree based on double-attribute and part-match(DAID3): First, if two best attributes have same or approximate entropy, DAID3 algorithm selects the two attributes to build the node, and the two attribute values and their combination as the node' branches. So a train instance can be covered one more times. Second, in the branch node, a new instance may be matched by one more branch path. In order to select the best one; we measure each branch node by the strong. Finally, for a new instance, if there is not a path from the root to a leaf to fully match the instance, DAID3 finds the longest part path, and returns the max-strong class label of the end note. So we labeled each branch node by the max-strong class label. Experimental results show that DAID3 algorithm has higher accuracy than the decision tree more.