朴素贝叶斯法 - 垃圾邮件分类

Peren ·
更新时间:2024-11-13
· 913 次阅读

本文基于朴素贝叶斯构建一个分类垃圾邮件的模型,研究对象是英文的垃圾邮件。

邮件内容保存在txt文件中,其中分为训练样本train和测试样本test。
在训练样本中正常邮件命名为:pos;垃圾邮件命名为:neg。
同时,可以将待分类的测试样本放入测试文件test中的对于pos,或者neg下,用来进行测试,如果是垃圾邮件则类别为0,反之类别为1。

在朴素贝叶斯法进行垃圾邮件的分类的思想中:有一个方法与一个假设:
  贝叶斯定理:求解p(c|x)的问题变成了求解p(x|c)的问题
  特征条件独立假设 :X的n个特征在某类确定的条件下都是条件独立的。
具体可以看:[监督学习] 朴素贝叶斯法.
在GitHub上的数据: 数据及代码.

1 邮件分类算法步骤

1 建立词汇表,(用来收集训练集中所有邮件的词汇。)
2 每封邮件词汇向量,(通过所有词汇来确定每封邮件中词汇出现的数量。)
3 计算先验概率p(b)
4 计算每个词汇的条件概率
5 用例测试

完整程序:

# -*- coding: utf-8 -*- # @Time : 2020/4/16 21:09 # @Author : Zudy ''' 1. 基于朴素贝叶斯的垃圾邮件分类 ''' from sklearn import datasets from time import time import numpy import re import os import random def load_data(folder_path): print("Loading dataset ...") loadTime = time() datalist = datasets.load_files(folder_path) #datalist是一个Bunch类,其中重要的数据项有 #data:原始数据 #filenames:每个文件的名称 #target:类别标签(子目录的文件从0开始标记了索引) #target_names:类别标签(子目录的具体名称) #输出总文档数和类别数 print("summary: {0} documents in {1} categories.".format(len(datalist.data),len(datalist.target_names))) #加载数据所用的时间 print("Load data in {0}seconds".format(time() - loadTime)) #去停用词操作 #datalist.data = [word for word in datalist.data if(word not in stopwords.words('english'))] return datalist # 返回集合列表 def word_create(ori_data): print("\nVectorzing dataset ...") word_dic = set([]) #建立一个集合列表 #词向量的时间 vectorTime = time() #词典的构造 for doc in ori_data.data: #doc是byte,这里将byte转化为string doc = str(doc, encoding = "utf-8") #使用正则表达式将特殊符号去除 doc = re.sub("[\s+\.\!\/_,$%^*(+\"\'-]+|[+——!,。?、~@#¥%……&*()]+", " ", doc) #使用默认的空格方式将email分隔开,然后转化为小写字母,与原集合取并集 word_dic = word_dic|set(doc.lower().split()) #向量化的时间和词典中词的数量 print("Vectorzing time:{0}\nThe number of word_dictionary:{1}".format(vectorTime,len(word_dic))) return list(word_dic) def doc_represent(wordDic,ori_data): #创建一个文档数(行)*词向量(列)长度的二维数组 doc_re = numpy.zeros((len(ori_data.data),len(wordDic)),dtype= numpy.int) #计数器 count = 0 #用来记录词向量表示时间 representTime = time() for doc in ori_data.data: #同word_create函数,进行同样的操作 doc = str(doc, encoding = "utf-8") doc = re.sub("[\s+\.\!\/_,$%^*(+\"\'-]+|[+——!,。?、~@#¥%……&*()]+", " ", doc) for word in doc.lower().split(): if word in wordDic: #将对应词向量位置置1 doc_re[count][wordDic.index(word)] = 1 count = count+1 print("Represent doc time:{0}\nThe number of doc:{1}".format(representTime-time(),len(doc_re))) #返回表示文档的二维数组 return doc_re def pre_probabilty(ori_data): s_pre_pro = [] #正常邮件的先验概率 P_normal = (normal + 1.0)/(len(ori_data.data) + 2.0) s_pre_pro.append(P_normal) #垃圾邮件的先验概率 P_spam = (spam + 1.0)/(len(ori_data.data) + 2.0) s_pre_pro.append(P_spam) #返回先验概率的列表 return s_pre_pro #计算每个词在正常邮件垃圾邮件中的数目 def wordNum_email(email_repre,wordDic): #用二维向量存储 num_word = numpy.zeros((2,len(wordDic)),dtype= numpy.int) for i in range(len(wordDic)): #在正常邮件的数目 for j in range(normal): num_word[0][i] += email_repre[j][i] #在垃圾邮件中的数目 for j in range(normal, spam+normal): num_word[1][i] += email_repre[j][i] return num_word #条件概率 def con_probabilty(email_repre,wordDic): #得到每个词汇在正常邮件、垃圾邮件中的数目 word_num = wordNum_email(email_repre,wordDic) word_pro = numpy.zeros((2,len(wordDic)),dtype = numpy.double) for i in range(len(wordDic)): word_pro[0][i] = round((word_num[0][i]+1)/(normal + 2),8) word_pro[1][i] = round((word_num[1][i]+1)/(spam + 2 ),8) return word_pro #得到每个类别中的文档数 def class_num(path,class_name): count = 0 path=path+"/"+class_name for root, dirs, files in os.walk(path): # 遍历统计 for each in files: count += 1 return count #测试 def test_spam(test_repre,pre_pro,con_pro): email_pro = numpy.zeros((len(test_repre),2),dtype = numpy.double) email_judge = [] normal_num = 0 spam_num = 0 for i in range(len(test_repre)): email_pro[i][0] = round(pre_pro[0],8) email_pro[i][1] = round(pre_pro[1],8) for j in range(len(test_repre[0])): if test_repre[i][j] != 0: email_pro[i][0] *= con_pro[0][j] email_pro[i][1] *= con_pro[1][j] if email_pro[i][0] > email_pro[i][1] : email_judge.append(0) elif email_pro[i][0] 0.5: email_judge.append(1) else: email_judge.append(0) for i in range(normal_test): if email_judge[i] == 0: normal_num +=1 for i in range(normal_test,len(test_repre)): if email_judge[i] == 1: spam_num +=1 print("email_judge=") print(email_judge) print("normal_num="+str(normal_num)+"\nspam_num="+str(spam_num)) return (normal_num + spam_num)/len(test_repre) if __name__ == "__main__": # 训练集和测试集的路径 train_path = "D:/Python/Python_learning/Book_code/LH_mechine_learning/bayes/spamDataset/email/train1" test_path = "D:/Python/Python_learning/Book_code/LH_mechine_learning/bayes/spamDataset/email/test1" train_list = load_data(train_path) test_list = load_data(test_path) normal = class_num(train_path,"pos") # 正常邮件的数目 spam = class_num(train_path,"neg") # 垃圾邮件的数目 WordDictionary = word_create(train_list) # 建立词汇表 docRepre = doc_represent(WordDictionary,train_list) # 将训练数据进行向量表示 prePro = pre_probabilty(train_list) conPro = con_probabilty(docRepre,WordDictionary) print("\npreProbablity:",prePro) # 计算先验概率 print("conProbablity:",conPro) # 计算条件概率 testRepre = doc_represent(WordDictionary,test_list) # 测试数据的向量表示 normal_test = class_num(test_path, "pos") # 正常邮件的数目 spam_test = class_num(test_path, "neg") # 垃圾邮件的数目 test_accuracy = test_spam(testRepre,prePro,conPro) # 测试数据的准确率 print ("test accuracy") print(test_accuracy) 2 测试结果 Loading dataset ... summary: 43 documents in 2 categories. Load data in 0.008994102478027344seconds Loading dataset ... summary: 36 documents in 2 categories. Load data in 0.007995843887329102seconds Vectorzing dataset ... Vectorzing time:1587043343.3496442 The number of word_dictionary:2426 Represent doc time:-0.5976784229278564 The number of doc:43 preProbablity: [0.5777777777777777, 0.4222222222222222] conProbablity: [[0.59259259 0.07407407 0.03703704 ... 0.11111111 0.07407407 0.03703704] [0.7 0.05 0.15 ... 0.05 0.05 0.1 ]] Represent doc time:-0.5137045383453369 The number of doc:36 email_judge= [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 1, 1, 0, 1, 0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1] normal_num=16 spam_num=13 test accuracy 0.8055555555555556

参考文献:
链接: 基于朴素贝叶斯的垃圾邮件分类.


作者:三个半_Z



朴素贝叶斯 分类 贝叶斯 邮件

需要 登录 后方可回复, 如果你还没有账号请 注册新账号