决策树
http://www.cnblogs.com/fengfenggirl/p/classsify_decision_tree.html
决策树是一种基本的分类和回归方法。优点是具有可读性,分类速度快,学习时,利用训练数据,根据损失函数最小化得原则建立决策树模型。
缺点是容易过拟合,对训练数据的分类能力好,泛化能力差。
决策树通常分为3个部分:特征选择,决策树的生成和决策树的修剪。
决策树的学习算法通常是地跪地选择一个最优的特征,并根据该特征对训练数据进行分割,使得各个子数据集具有较好的分类过程,这个过程对应
着特征空间的划分,也对应这决策树的构建,开始,构建根结点,所有数据都放在根结点。
选择一个最优特征,按照这一特征将数据集划分为多个子集,这些子集有一个在当前条件下最好的分类。
如果这些子集已经能够被正确的分类,就构造叶子结点,并将这些子集放到对应的叶子结点中;
如果还有子集不能够被正确分类,就对这些子集选择最优特征,并继续进行分割。
如此递归进行下去,直到所有训练子集都能够被正确分类,或者没有合适的特征。
最后每个子集都被正确的分到叶子节点上,即有了明确的类。
决策树主要有ID3,C4.5,和CART算法。
ID3和C4.5的区别是特征划分的标准,一个是信息增益,一个是用信息增益比。
信息增益可以表示:
\(g(D,A)=H(D)-H(D|A)\) ,其中H()表示熵,熵越大代表不确定性越大。
这个公式的意义就在于,使用特征A进行分类后,整体的不确定性会大大减小。
但是这个公式是有个问题,它计算的是绝对值,如果在某些极端情况下,特征A有可以将每个样本变为树的节点,那H(D|A)就为0,
这时候信息增益最大,但这个并不是并不是我们想要的特征。同时如果一个特征值的经验熵大的时候,信息增益会偏大,这个时候,
使用信息增益比,会对该这种情况进行校正,这就是C4.5算法。
看到书上信息增益比的公式是:
我觉得这里的H(D)应该是H(A)。
分类与回归树(CART)算法假设决策树是一个二叉树,内部节点特征的取值为“是”和“否”,左边选取是的分支,右边选取否的分支。
这样的决策树等价于递归地二分每个特征,并将输入空间特征有效地划分为有限个单元,并在这些单元上确定的预测概率分布。
CART使用基尼指数进行计算
ID3算法实现:
# -*- coding: utf-8 -*- import math import operator import numpy as np class Id3(object): """ 决策树ID3算法实现 """ def __init__(self): pass def cal_entropy(self, data_set): """ 计算熵 :param data_set: [1, 1, 1, 0, 0] :return: """ num_entries = len(data_set) featrue_dict = {} for feature in data_set: lable = feature[-1] if lable in featrue_dict.keys(): featrue_dict[lable] += 1 else: featrue_dict[lable] = 1 entropy = 0.0 for key in featrue_dict.keys(): prob = float(featrue_dict[key]) / num_entries if prob != 0: entropy -= prob * math.log(prob, 2) return entropy def get_sub_data(self, data_set, feat_index, value): """ 获取用来计算条件熵的部分数据 :param data_set: :return: """ sub_data = [] num_entries = len(data_set) for i in xrange(num_entries): if data_set[i][feat_index] == value: sub_data.append([data_set[i][feat_index], data_set[i][-1]]) return sub_data def get_max_info_gain_ratio_and_fature(self, data_set): """ 得到最大增益信息和对应的 feature :param data_set: :return: max_info_gain , feature """ #初始化数据 num_features = len(data_set[0]) - 1 base_entropy = self.cal_entropy(data_set) max_info = 0.0 max_feature = -1 #计算每个特征的熵 for i in xrange(num_features): feature = [sample[i] for sample in data_set] unique_feature_values = set(feature) tmp_entropy = 0.0 for value in unique_feature_values: sub_data = self.get_sub_data(data_set, i, value) prob = len(sub_data) / float(len(data_set)) tmp_entropy += prob * self.cal_entropy(sub_data) #保留最大增益信息和特征 info_gain = base_entropy - tmp_entropy #info_gain_ratio = (base_entropy - tmp_entropy) / base_entropy if info_gain > max_info: max_info = info_gain max_feature = i return max_info, max_feature def get_majority_cnt(self, classlist): """ 获取数据集中实例数最大的类的名称 :param classlist: :return: """ class_count = {} for data in classlist: if data in class_count.keys(): class_count[data] += 1 else: class_count[data] = 0 reverse_class_count = sorted(class_count.iteritems(), key=operator.itemgetter(1), reverse=True) return reverse_class_count[0][0] def split_data_set(self, data_set, feature_index, value): """ 去掉特征集合中对应的feather :param feature_index: :param value: :return: """ res_data = [] for sample in data_set: if sample[feature_index] == value: split_sample = sample[:feature_index] split_sample.extend(sample[feature_index+1:]) res_data.append(split_sample) return res_data def creata_decision_tree(self, data_set , labels , threshold): """ :param data_set: 数据集 :param labels: 标签集 :param threshold: 阈值 :return: 返回一棵树 """ class_list = [example[-1] for example in data_set] #如果只有一种结果了,那就停止 if class_list.count(class_list[0]) == len(class_list): return class_list[0] if len(data_set[0]) == 1: return self.get_majority_cnt(class_list) max_ratio, best_feature_index = self.get_max_info_gain_ratio_and_fature(data_set) feature_list = [example[best_feature_index] for example in data_set] if max_ratio < threshold: return self.get_majority_cnt(feature_list) best_feature_label = labels[best_feature_index] my_tree = {best_feature_label: {}} del(labels[best_feature_index]) unique_values = set(feature_list) for value in unique_values: sub_labels = labels[:] my_tree[best_feature_label][value] = self.creata_decision_tree(self.split_data_set(data_set, best_feature_index, value), sub_labels, threshold) return my_tree def classify(self, inputTree, featLabels, testVec): firstStr = inputTree.keys()[0] secondDict = inputTree[firstStr] featIndex = featLabels.index(firstStr) for key in secondDict.keys(): if testVec[featIndex] == key: if type(secondDict[key]).__name__ == 'dict': classLabel = self.classify(secondDict[key], featLabels, testVec) else: classLabel = secondDict[key] return classLabel if __name__ == '__main__': obj = Id3() data_set = [[1,1,'yes'], [1,1, 'yes'], [1,0,'no'], [0,1,'no'], [0,1,'no']] labels = ['no surfacing','flippers'] create_labels = labels[:] tree = obj.creata_decision_tree(data_set, create_labels, 0.1) print obj.classify(tree, labels, [1, 0])
CART算法实现:
# -*- coding: utf-8 -*- """ classfication and regression tree 分类与回归树 """ import operator class Node(object): def __init__(self, feat, val, left=None, right=None, result=None): self.feature = feat self.split_value = val self.left_node = left self.right_node = right self.result = result class Cart(object): """ 分类与回归树, 假设决策树为二叉树 """ def __init__(self): pass def cal_gini(self, data_set, num_samples): """ 计算基尼指数, 1 - sum(pi) :param data_set: :return: """ num_entries = num_samples featrue_dict = {} for feature in data_set: lable = feature[-1] if lable in featrue_dict.keys(): featrue_dict[lable] += 1 else: featrue_dict[lable] = 1 temp = 0.0 for key in featrue_dict.keys(): prob = float(featrue_dict[key]) / num_entries temp += pow(prob, 2) return 1 - temp def get_sub_data(self, data_set, feat_index, value): """ 获取用来计算条件熵的部分数据 :param data_set: :return: """ sub_data = [] num_entries = len(data_set) for i in xrange(num_entries): if data_set[i][feat_index] == value: sub_data.append([data_set[i][feat_index], data_set[i][-1]]) return sub_data def get_best_feature_and_split(self, data_set): """ 遍历每一个特征A,对其可能取值a,计算A=a的时候的基尼指数 :param data_set: :return: """ #样本个数 num_samples = len(data_set) #特征个数 num_features = len(data_set[0]) - 1 #最小GINI和最佳feature min_gini = 9999.0 best_feature_index = -1 split_val = None for i in xrange(0, num_features): #计算特征值的每个变量的GINI feature = [example[i] for example in data_set] unique_feature = set(feature) for val in unique_feature: data_include_val, data_exclude_val = self.split_data(data_set, i, val) gini_data_include_val = self.cal_gini(data_include_val, num_samples) gini_data_include_val *= float(len(data_include_val)) / num_samples gini_data_exclude_val = self.cal_gini(data_exclude_val, num_samples) gini_data_exclude_val *= float(len(data_exclude_val)) / num_samples gini = gini_data_include_val + gini_data_exclude_val if gini < min_gini: min_gini = gini best_feature_index = i split_val = val return best_feature_index, split_val def split_data(self, data_set, feature_index, split_value): """ 分割数据,对于每个特征的取值a,分为两类,一类特征取值为a,一类特征取值不为a :param data_set: :param feature_index: 特征值下标 :param split_value: 分割的值 :return: """ #特征值为a的list data_include_a = [] #特征值不为a的list data_exclude_a = [] for example in data_set: tmp_list = example[:feature_index] tmp_list.extend(example[feature_index + 1:]) if example[feature_index] == split_value: data_include_a.append(tmp_list) else: data_exclude_a.append(tmp_list) return data_include_a, data_exclude_a def get_majority_cnt(self, classlist): """ 获取数据集中实例数最大的类的名称 :param classlist: :return: """ class_count = {} for data in classlist: if data in class_count.keys(): class_count[data] += 1 else: class_count[data] = 0 reverse_class_count = sorted(class_count.iteritems(), key=operator.itemgetter(1), reverse=True) return reverse_class_count[0][0] def build_cart_tree(self, data_set): """ 创建树 :param data_set: :return: """ class_list = [example[-1] for example in data_set] #如果只有一种结果了,那就停止 if class_list.count(class_list[0]) == len(class_list): return Node(None, None, result=class_list[0]) feature_index, split_val = self.get_best_feature_and_split(data_set) m1, m2 = self.split_data(data_set, feature_index, split_val) root = Node(feature_index, split_val) root.left_node = self.build_cart_tree(m1) root.right_node = self.build_cart_tree(m2) return root def cart_classfy(self, root, test_vec): if root.feature is None: return root.result if test_vec[root.feature] == root.split_value: return self.cart_classfy(root.left_node, test_vec) else: return self.cart_classfy(root.right_node, test_vec) if __name__ == '__main__': obj = Cart() data_set = [[1,1,'yes'], [1,1, 'yes'], [1,0,'no'], [0,1,'no'], [0,1,'no']] labels = ['no surfacing','flippers'] create_labels = labels[:] tree = obj.build_cart_tree(data_set) print obj.cart_classfy(tree, [1, 0])
posted on 2016-03-08 20:05 walkwalkwalk 阅读(333) 评论(0) 编辑 收藏 举报