英语论文网

Under normal circumstances , or with a larger probability to say that the smaller the tree , the stronger the predictive ability of the tree . To construct a decision tree as small as possible , the key is attribute selection measure . It can only be taken to choose a good heuristic strategy logic judgment or attributes. Property right choice depends on a variety of examples subset impurity metrics . Not purity metrics , including information gain, information gain ratio , Gini index , distance metric , J-measure, G statistics , statistics, weight of evidence , the minimum description length (MLP), orthogonal method, correlation and Relief. Different measures have different effects , especially for multi-valued attributes . [ 7 ]

当决策树创建时，由于数据中的噪声和孤立点，许多分支反映的是训练集数据中的异常，为了处理这种过分适应数据问题，采用树剪枝方法，剪去决策树中的最不可靠分支，减少决策树的尺寸，这样可以达到较快的分类，提高决策树独立于数据正确分类的能力。下面介绍2种常用的树剪枝方法：预剪枝和后剪枝。

When a tree is created, because the data in the noise and isolated points , many branches reflect anomalies in the data training set , in order to deal with this problem too accommodate data , using the tree pruning method , cut tree in the least reliability branch , reducing the size of the tree , so that the classification can be achieved quickly , independent of the data to improve the correct classification tree capacity . Here are two kinds of commonly used tree pruning method : Pre- pruning and post- pruning .

预剪枝方法是在完全正确分类训练集之前，通过一定的法则较早地停止树的构造而对树进行“剪枝”，一旦停止，节点成为树叶，该树叶可能持有子集样本中最频繁的类，或这些样本的概率分布。

Pre- pruning method is entirely before the training set correctly classified by certain rules to stop earlier tree structure and the tree " pruning ", once stopped , the nodes become leaves, the leaves may hold a subset of samples of the most frequent class , or the probability distribution of these samples . In the tree structure , the use of statistics , and other information gain measure , evaluate the quality of the split , if the nodes into a sample will result in less than a predetermined threshold, split , the given subset further divided stops. This algorithm is relatively simple , high efficiency, suitable for large-scale problems . [ 8 ]

后剪枝方法由“完全生长”的树剪去不可靠分枝，通过删除节点的分枝，剪掉树节点。它利用训练集样本或剪枝样本数据，检验决策子树对目标变量的预测精度，并计算出相应的错误率。

After pruning method from " fully grown " tree branch cut is not reliable , by removing nodes branching, cut tree nodes. It uses the training set or pruning sample data , inspection decisions subtree prediction accuracy of the target variable , and calculate the corresponding error rate. Users can pre-set a maximum allowable error rate. When pruning reaches a certain depth, the error rate is higher than the calculated maximum allowable error rate , then stop pruning, otherwise continue pruning . Finally got one with the smallest expected error rate of the decision tree . [ 8 ]

Decision tree algorithm has good scalability refers algorithm capable of handling large amounts of data or the ability to accelerate the data mining process , from the large amounts of data quickly and accurately focused discover hidden in one of the main classification rules . Scalability, including the strengthening of the decision tree algorithm , and scalability . Study the scalability of decision tree algorithm main reasons are the foll