摘要
神经网络集成方法具有比单个神经网络更强的泛化能力,却因为其黑箱性而难以理解;决策树算法因为分类结果显示为树型结构而具有良好的可理解性,泛化能力却比不上神经网络集成。该文将这两种算法相结合,提出一种决策树的构造算法:使用神经网络集成来预处理训练样本,使用C4.5算法处理预处理后的样本并生成决策树。该文在UC I数据上比较了神经网络集成方法、决策树C4.5算法和该文算法,实验表明:该算法具有神经网络集成方法的强泛化能力的优点,其泛化能力明显优于C4.5算法;该算法的最终结果显示为决策树,显然具有良好的可理解性。
Neural network ensemble is with stronger generalization ability compared with a single neural network. But the ensemble is lack of comprehensibility because it is regarded as a ‘black box＇. And decision tree is with good comprehensibility. But its generalization ability can not be compared with neural network ensemble. In this paper, an algorithm for building a decision tree is proposed which combines the merits of both the neural network ensemble and the decision tree. The algorithm uses neural network ensemble to reprocess the training set then forms a C4. 5 decision tree. Experimental results are compared among neural network ensemble, decision tree and the algorithm introduced in this paper. Experiments show that the algorithm in this paper is with strong generalization ability inherited from neural network ensemble and it has stronger generalization than C4. 5 decision tree. Because the result of the algorithm is shown as a tree, the algorithm has good comprehensibility.
出处
《计算机仿真》
CSCD
2006年第11期95-98,共4页
Computer Simulation
关键词
神经网络集成
神经网络分类器
决策树
Neural network ensemble
Neural network classifiers
Decision tree
