乍看起来似乎是要求一个概率,还要先得到额外三个概率,有用么?其实这个简单的公式非常贴切人类推理的逻辑,即通过可以观测的数据,推测不可观测的数据。举个例子,也许你在办公室内不知道外面天气是晴天雨天,但是你观测到有同事带了雨伞,那么可以推断外面八成在下雨。
若X 是要输入的随机变量,则Y 是要输出的目标类别。对X 进行分类,即使求的使P(Y|X) ***的Y值。若X 为n 维特征变量 X = {A1, A2, …..An} ,若输出类别集合为Y = {C1, C2, …. Cm} 。
X 所属最有可能类别 y = argmax P(Y|X), 进行如下推导:
朴素贝叶斯的学习
有公式可知,欲求分类结果,须知如下变量:
各个类别的条件概率,
输入随机变量的特质值的条件概率
示例代码:
- import copy
- class native_bayes_t:
- def __init__(self, character_vec_, class_vec_):
- """
- 构造的时候需要传入特征向量的值,以数组方式传入
- 参数1 character_vec_ 格式为 [("character_name",["","",""])]
- 参数2 为包含所有类别的数组 格式为["class_X", "class_Y"]
- """
- self.class_set = {}
- # 记录该类别下各个特征值的条件概率
- character_condition_per = {}
- for character_name in character_vec_:
- character_condition_per[character_name[0]] = {}
- for character_value in character_name[1]:
- character_condition_per[character_name[0]][character_value] = {
- 'num' : 0, # 记录该类别下该特征值在训练样本中的数量,
- 'condition_per' : 0.0 # 记录该类别下各个特征值的条件概率
- }
- for class_name in class_vec:
- self.class_set[class_name] = {
- 'num' : 0, # 记录该类别在训练样本中的数量,
- 'class_per' : 0.0, # 记录该类别在训练样本中的先验概率,
- 'character_condition_per' : copy.deepcopy(character_condition_per),
- }
- #print("init", character_vec_, self.class_set) #for debug
- def learn(self, sample_):
- """
- learn 参数为训练的样本,格式为
- [
- {
- 'character' : {'character_A':'A1'}, #特征向量
- 'class_name' : 'class_X' #类别名称
- }
- ]
- """
- for each_sample in sample:
- character_vec = each_sample['character']
- class_name = each_sample['class_name']
- data_for_class = self.class_set[class_name]
- data_for_class['num'] += 1
- # 各个特质值数量加1
- for character_name in character_vec:
- character_value = character_vec[character_name]
- data_for_character = data_for_class['character_condition_per'][character_name][character_value]
- data_for_character['num'] += 1
- # 数量计算完毕, 计算最终的概率值
- sample_num = len(sample)
- for each_sample in sample:
- character_vec = each_sample['character']
- class_name = each_sample['class_name']
- data_for_class = self.class_set[class_name]
- # 计算类别的先验概率
- data_for_class['class_per'] = float(data_for_class['num']) / sample_num
- # 各个特质值的条件概率
- for character_name in character_vec:
- character_value = character_vec[character_name]
- data_for_character = data_for_class['character_condition_per'][character_name][character_value]
- data_for_character['condition_per'] = float(data_for_character['num']) / data_for_class['num']
- from pprint import pprint
- pprint(self.class_set) #for debug
- def classify(self, input_):
- """
- 对输入进行分类,输入input的格式为
- {
- "character_A":"A1",
- "character_B":"B3",
- }
- """
- best_class = ''
- max_per = 0.0
- for class_name in self.class_set:
- class_data = self.class_set[class_name]
- per = class_data['class_per']
- # 计算各个特征值条件概率的乘积
- for character_name in input_:
- character_per_data = class_data['character_condition_per'][character_name]
- per = per * character_per_data[input_[character_name]]['condition_per']
- print(class_name, per)
- if per >= max_per:
- best_class = class_name
- return best_class
- character_vec = [("character_A",["A1","A2","A3"]), ("character_B",["B1","B2","B3"])]
- class_vec = ["class_X", "class_Y"]
- bayes = native_bayes_t(character_vec, class_vec)
- sample = [
- {
- 'character' : {'character_A':'A1', 'character_B':'B1'}, #特征向量
- 'class_name' : 'class_X' #类别名称
- },
- {
- 'character' : {'character_A':'A3', 'character_B':'B1'}, #特征向量
- 'class_name' : 'class_X' #类别名称
- },
- {
- 'character' : {'character_A':'A3', 'character_B':'B3'}, #特征向量
- 'class_name' : 'class_X' #类别名称
- },
- {
- 'character' : {'character_A':'A2', 'character_B':'B2'}, #特征向量
- 'class_name' : 'class_X' #类别名称
- },
- {
- 'character' : {'character_A':'A2', 'character_B':'B2'}, #特征向量
- 'class_name' : 'class_Y' #类别名称
- },
- {
- 'character' : {'character_A':'A3', 'character_B':'B1'}, #特征向量
- 'class_name' : 'class_Y' #类别名称
- },
- {
- 'character' : {'character_A':'A1', 'character_B':'B3'}, #特征向量
- 'class_name' : 'class_Y' #类别名称
- },
- {
- 'character' : {'character_A':'A1', 'character_B':'B3'}, #特征向量
- 'class_name' : 'class_Y' #类别名称
- },
- ]
- input_data ={
- "character_A":"A1",
- "character_B":"B3",
- }
- bayes.learn(sample)
- print(bayes.classify(input_data))
总结:
朴素贝叶斯分类实现简单,预测的效率较高
朴素贝叶斯成立的假设是个特征向量各个属性条件独立,建模的时候需要特别注意
原文链接:http://www.cnblogs.com/zhiranok/archive/2012/09/22/native_bayes.html
【编辑推荐】