贝叶斯理论
我们有一堆带标记的样本(包含 特征 和 类别),可以从中统计得到 p(特征|类别)
根据
全概率公式:
P(Y,X)=P(Y|X)P(X)=P(X|Y)P(Y)
得到
贝叶斯公式:
P(Y|X)=P(X|Y)P(Y)P(X)
从机器学习的视角,
X
认为是 具有某特征 ,
P(“属于某类”|“具有某特征”)=P(“具有某特征”|“属于某类”)P(“属于某类”)P(“具有某特征”)
独立假设
实际上,具有的特征往往是多维的
P(features|class)=P(f0,f1,...,fn|c)
如果是2维,可以写成
P(f0,f1|c)=P(f0|c)P(f1|c,f0)
加上条件独立假设,认为特征之间是相互独立的,则
P(f0,f1|c)=P(f0|c)P(f1|c)
即:
P(f0,f1,…,fn|c)=ΠniP(fi|c)
贝叶斯分类器
回到机器学习,其实我们就是根据样本预料,对每个类别计算一个概率
P(ci)
,然后再计算所有特征的条件概率
P(fi|ci)
,那么分类的时候就是根据朴素贝叶斯方法选择一个概率最大的类别:
P(ci|f0,f1,…,fn)=P(ci)P(f0,f1,…,fn)ΠnjP(fj|ci)
文本分类问题
采用朴素贝叶斯进行新闻类别分类
数据集
9类新闻数据,每类有10篇新闻文本
C000008 财经
C000010 IT
C000013 健康
C000014 体育
C000016 旅游
C000020 教育
C000022 招聘
C000023 文化
C000024 军事
Python3.6代码
import os
import random
import jieba #处理中文
#import nltk #处理英文
from sklearn.naive_bayes import MultinomialNB
import matplotlib.pyplot as plt
# 文本处理,也就是样本生成过程
def text_processing(folder_path, test_size=0.2):
folder_list = os.listdir(folder_path)
data_list = []
class_list = []
# 遍历文件夹
for folder in folder_list:
new_folder_path = os.path.join(folder_path, folder)
files = os.listdir(new_folder_path)
# 读取文件
j = 1
for file in files:
if j > 100: # 怕内存爆掉,只取100个样本文件,你可以注释掉取完
break
with open(os.path.join(new_folder_path, file), 'rb') as fp:
text = fp.read()
## 是的,随处可见的jieba中文分词
# jieba.enable_parallel(4) # 开启并行分词模式,参数为并行进程数,不支持windows
word_cut = jieba.cut(text, cut_all=False) # 精确模式,返回的结构是一个可迭代的genertor
word_list = list(word_cut) # genertor转化为list,每个词unicode格式
# jieba.disable_parallel() # 关闭并行分词模式
data_list.append(word_list) # 训练集list #list里套list,每个list是没篇文章分完词的Unicode编码
class_list.append(folder) # 类别,即文件夹的名字(代表类别)
j += 1
## 粗暴地划分训练集和测试集
data_class_list = list(zip(data_list, class_list)) # 将列表对应元素合并,创建一个元组队的列表
random.shuffle(data_class_list)
index = int(len(data_class_list) * test_size) + 1 # 20% 当做训练集,index为19
train_list = data_class_list[index:] # 19-89共71个设为训练集
test_list = data_class_list[:index] # 0-18共19个设为测试集
train_data_list, train_class_list = zip(*train_list) # 将映射的元组unzip成原来的两个list
test_data_list, test_class_list = zip(*test_list)
# 其实可以用sklearn自带的部分做
# train_data_list, test_data_list, train_class_list, test_class_list = sklearn.cross_validation.train_test_split(data_list, class_list, test_size=test_size)
# 统计词频放入all_words_dict
all_words_dict = {} # 字典类型
for word_list in train_data_list:
for word in word_list: # 遍历训练集中每个词语
if word in all_words_dict:
all_words_dict[word] += 1
else:
all_words_dict[word] = 1
# key函数利用词频进行降序排序
all_words_tuple_list = sorted(all_words_dict.items(), key=lambda f: f[1], reverse=True) # 内建函数sorted参数需为list
all_words_list = list(zip(*all_words_tuple_list))[0] # 将元组分成两个list,只要第一个list,all_words_list为训练集中所有词语按词频排序的list
return all_words_list, train_data_list, test_data_list, train_class_list, test_class_list
#粗暴的词统计
def make_word_set(words_file):
words_set = set()#set类型没有重复的元素
with open(words_file, 'rb') as fp:
for line in fp.readlines():#按行读,每行是一个停用词
word = line.strip().decode("utf-8")#去除首位空格
if len(word)>0 and word not in words_set: # 去重
words_set.add(word)
return words_set
# 选取特征词
#去掉停用词,和词频高的(比较normal)的词,得到说有的特征词集,构成词袋
def words_dict(all_words_list, deleteN, stopwords_set=set()):
feature_words = []
n = 1
for t in range(deleteN, len(all_words_list), 1):#去掉词频较高的前deleteN个词,可能是停用词或比较normal的词
if n > 1000: # feature_words的维度1000
break
if not all_words_list[t].isdigit() and all_words_list[t] not in stopwords_set and 1<len(all_words_list[t])<5:
feature_words.append(all_words_list[t])
n += 1
return feature_words
# 文本特征
def text_features(train_data_list, test_data_list, feature_words, flag='nltk'):
def text_features(text, feature_words):
text_words = set(text)
## -----------------------------------------------------------------------------------
if flag == 'nltk':
## nltk特征 dict
features = {word:1 if word in text_words else 0 for word in feature_words}
elif flag == 'sklearn':
## sklearn特征 list
features = [1 if word in text_words else 0 for word in feature_words]#特征词汇中,如果出现在文本中,为1 ,否则为0,features的长度等于feature_words
else:
features = []
## -----------------------------------------------------------------------------------
return features
train_feature_list = [text_features(text, feature_words) for text in train_data_list]
test_feature_list = [text_features(text, feature_words) for text in test_data_list]
return train_feature_list, test_feature_list
# 分类,同时输出准确率等
def text_classifier(train_feature_list, test_feature_list, train_class_list, test_class_list, flag='nltk'):
## -----------------------------------------------------------------------------------
if flag == 'nltk':
## 使用nltk分类器
train_flist = zip(train_feature_list, train_class_list)
test_flist = zip(test_feature_list, test_class_list)
classifier = nltk.classify.NaiveBayesClassifier.train(train_flist)
test_accuracy = nltk.classify.accuracy(classifier, test_flist)
elif flag == 'sklearn':
## sklearn分类器
classifier = MultinomialNB().fit(train_feature_list, train_class_list)
test_accuracy = classifier.score(test_feature_list, test_class_list)
else:
test_accuracy = []
return test_accuracy
print ("start")
## 文本预处理
#将数据分为训练集和测试集,并得到训练集中所有词语的list,按词频排好序
folder_path = './Database/SogouC/Sample'
all_words_list, train_data_list, test_data_list, train_class_list, test_class_list = text_processing(folder_path, test_size=0.2)
# 生成stopwords_set
#从设置的文件中得到停用词集合
stopwords_file = './stopwords_cn.txt'
stopwords_set = make_word_set(stopwords_file)
## 文本特征提取和分类
# flag = 'nltk'
flag = 'sklearn'
deleteNs = range(0, 1000, 20) #[0,20,40,60,...]
test_accuracy_list = []
for deleteN in deleteNs:
# feature_words = words_dict(all_words_list, deleteN)
feature_words = words_dict(all_words_list, deleteN, stopwords_set)#得到所有特征词集(形成词袋)
train_feature_list, test_feature_list = text_features(train_data_list, test_data_list, feature_words, flag)#根据词袋得到对应的特征表示
test_accuracy = text_classifier(train_feature_list, test_feature_list, train_class_list, test_class_list, flag)## 分类,同时得到准确率等
test_accuracy_list.append(test_accuracy)
print (test_accuracy_list)
# 结果评价
#plt.figure()
plt.plot(deleteNs, test_accuracy_list)
plt.title('Relationship of deleteNs and test_accuracy')
plt.xlabel('deleteNs')
plt.ylabel('test_accuracy')
plt.show()
#plt.savefig('result.png')
print ("finished")
[0.57894736842105265, 0.52631578947368418, 0.52631578947368418, 0.57894736842105265, 0.57894736842105265, 0.52631578947368418, 0.52631578947368418, 0.52631578947368418, 0.52631578947368418, 0.52631578947368418, 0.52631578947368418, 0.52631578947368418, 0.63157894736842102, 0.63157894736842102, 0.63157894736842102, 0.63157894736842102, 0.63157894736842102, 0.57894736842105265, 0.57894736842105265, 0.63157894736842102, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265, 0.57894736842105265]
发现每次运行的结果不一样