摘要
神经元网络在多数情况下获得的精度要比决策树和回归算法精度高 ,这是因为它能适应更复杂的模型 .同时由于决策树通常每次只使用一个变量来分支 ,它所对应的识别空间只能是超矩形 ,这也就比神经元网络简单 ,精度不能与神经元网络相比 .然而神经元网络需要相对多的学习时间 ,并且其模型的可理解性不如决策树、Naive- Bayes等方法直观 .本文在进行两种算法对复杂模型的识别对比后 ,提出一个新的算法 NNTree,这是一个决策树和神经元网络杂交的算法 :决策树节点包含单变量的分支就象正常的决策树 ,但是叶子节点包含神经元网络分类器 .这个方法针对决策树处理大型数据的效能 ,保留了决策树的可理解性 ,改善了神经元网络的学习性能 ,同时可使这个分类器的精度大大超过这两种算法 。
Neural Networks can achive higher accuracy than decision tree and regression in most cases, because they are more flexible to the more complex model. At the same time, as decision tree use a variable to split each time, its recognised space is only a super rectangle, which means it is simpler than neural network and its accuracy is lower. But neural network needs much more time to train and it is harder to understand than decision tree. After compare these two algorithms, we propose a new algorithm, NNTree, which is a hybrid algorithm of decision tree and neural network: the decision tree nodes contain univariate splits as regular decision trees, but the leaves contain neural network classifiers. Based on that decision tree can scale up the accuracy to huge data, this approach retains the comprehension of decision tree and improves the performance of neural network, while its accarcy is higher than those two algorithms especially in larger database and complexer model.
出处
《小型微型计算机系统》
CSCD
北大核心
2001年第8期964-966,共3页
Journal of Chinese Computer Systems
