日志
机器学习与R之朴素贝叶斯分类器
朴素贝叶斯
1联合概率分布
p(x,y)=p(y)P(x|y) 或者p(A交B)=p(A)xp(B) p(A交B)不容易求,假设条件独立拆分成两个事件的乘积
2基本假设条件独立性
3利用贝叶斯定理 p(y|x)=P(x,y)/p(x)=p(y)P(x|y)/sum(y-i)[p(y)P(x|y)]
y=max p(y)P(x|y)
贝叶斯决策理论要求计算两个概率p1(x，y),p2(x, y):
如果p1(x，y) > p2 (x, y) , 那么属于类别1
如果p2(x, y) > pl(x, y) , 那么属于类别2
拉普拉斯估计--防止概率值为0
每一个似然函数 分子+1对分母加上分子中加上1的总数
在朴素贝叶斯使用数值特征采用数值特征离散化,找见数据分布分割点切分
1分割词-去大小写-去字母？-去停用词-去多余空格符合
2统计每个词在每条短信中出现的频率
3创建频率矩阵 行是短信中词是否出现出现可以是yes no 列是每个词 label可以是字符串
library(e1071)
sms_classifier <- naiveBayes(sms_train, sms_train_labels)
sms_test_pred <- predict(sms_classifier, sms_test)
例子 iris[,-5]不含第5项
data(iris)
m <- naiveBayes(iris[,-5], iris[,5])
m
table(predict(m, iris), iris[,5])
另一个朴素贝叶斯包 klaR NaiveBayes()
朴素贝叶斯分类器通常有两种实现方式：一种基于贝努利模型实现， 一种基于多项式模型实现
这里采用前一种实现方式。该实现方式中并不考虑词在文档中出现的次数， 只考虑出不出现，
因此在这个意义上相当于假设词是等权重的
垃圾短信识别
# read the sms data into the sms data frame
sms_raw <- read.csv("sms_spam.csv", stringsAsFactors = FALSE)
# examine the structure of the sms data 垃圾邮件标记为spam非垃圾ham 结构type+text
str(sms_raw)
# convert spam/ham to factor.字符串分类标签转换成因子比较好
sms_raw$type <- factor(sms_raw$type)
# examine the type variable more carefully
str(sms_raw$type)
table(sms_raw$type)
# build a corpus using the text mining (tm) package
#tm文本挖掘包
library(tm)
sms_corpus <- VCorpus(VectorSource(sms_raw$text))#创建语料库 VCorpus存储R文本文档
# examine the sms corpus
print(sms_corpus)
inspect(sms_corpus[1:2]) #查看1-2个语料库内容
as.character(sms_corpus[[1]])
lapply(sms_corpus[1:2], as.character)
# clean up the corpus using tm_map()字母转换成小写
#sms_corpus_clean <- tm_map(sms_corpus, content_transformer(tolower))
# show the difference between sms_corpus and corpus_clean
#as.character(sms_corpus[[1]])
#as.character(sms_corpus_clean[[1]])
sms_corpus_clean <- tm_map(sms_corpus, removeNumbers) # remove numbers去掉数字
sms_corpus_clean <- tm_map(sms_corpus_clean, content_transformer(tolower)) #字母转换成小写
sms_corpus_clean <- tm_map(sms_corpus_clean, removeWords, stopwords()) # remove stop words去掉停用词
sms_corpus_clean <- tm_map(sms_corpus_clean, removePunctuation) # remove punctuation去掉标点
# tip: create a custom function to replace (rather than remove) punctuation
removePunctuation("hello...world")
replacePunctuation <- function(x) { gsub("[[:punct:]]+", " ", x) }
replacePunctuation("hello...world")
# illustration of word stemming
library(SnowballC)
wordStem(c("learn", "learned", "learning", "learns"))
sms_corpus_clean <- tm_map(sms_corpus_clean, stemDocument)
sms_corpus_clean <- tm_map(sms_corpus_clean, stripWhitespace) # eliminate unneeded whitespace去掉多余的空格
# examine the final clean corpus
lapply(sms_corpus[1:3], as.character)
lapply(sms_corpus_clean[1:3], as.character)
# create a document-term sparse matrix创建一个稀疏矩阵
sms_dtm <- DocumentTermMatrix(sms_corpus_clean)
# alternative solution: create a document-term sparse matrix directly from the SMS corpus
#直接从语料库创建一个稀疏矩阵
sms_dtm2 <- DocumentTermMatrix(sms_corpus, control = list(
tolower = TRUE,
removeNumbers = TRUE,
stopwords = TRUE,
removePunctuation = TRUE,
stemming = TRUE
))
# alternative solution: using custom stop words function ensures identical result
# 使用自定义的停用词
sms_dtm3 <- DocumentTermMatrix(sms_corpus, control = list(
tolower = TRUE,
removeNumbers = TRUE,
stopwords = function(x) { removeWords(x, stopwords()) },
removePunctuation = TRUE,
stemming = TRUE
))
# compare the result
sms_dtm
sms_dtm2
sms_dtm3
# creating training and test datasets
sms_dtm_train <- sms_dtm[1:4169, ]
sms_dtm_test <- sms_dtm[4170:5559, ]
# also save the labels
sms_train_labels <- sms_raw[1:4169, ]$type
sms_test_labels <- sms_raw[4170:5559, ]$type
# check that the proportion of spam is similar查看类别比例
prop.table(table(sms_train_labels))
prop.table(table(sms_test_labels))
# word cloud visualization词云可视化
library(wordcloud)
#从语料库直接创建词云
wordcloud(sms_corpus_clean, min.freq = 50, random.order = FALSE)
# subset the training data into spam and ham groups
spam <- subset(sms_raw, type == "spam")
ham <- subset(sms_raw, type == "ham")
wordcloud(spam$text, max.words = 40, scale = c(3, 0.5))
wordcloud(ham$text, max.words = 40, scale = c(3, 0.5))
sms_dtm_freq_train <- removeSparseTerms(sms_dtm_train, 0.999)
sms_dtm_freq_train
# indicator features for frequent words找出不少于5条短信的单词--减少特征
findFreqTerms(sms_dtm_train, 5)
# save frequently-appearing terms to a character vector
sms_freq_words <- findFreqTerms(sms_dtm_train, 5)
str(sms_freq_words)
# create DTMs with only the frequent terms
sms_dtm_freq_train <- sms_dtm_train[ , sms_freq_words]
sms_dtm_freq_test <- sms_dtm_test[ , sms_freq_words]
# convert counts to a factor转为成yes no
convert_counts <- function(x) {
x <- ifelse(x > 0, "Yes", "No")
}
# apply() convert_counts() to columns of train/test data MARGIN = 2 2是列1是行
sms_train <- apply(sms_dtm_freq_train, MARGIN = 2, convert_counts)
sms_test <- apply(sms_dtm_freq_test, MARGIN = 2, convert_counts)
## Step 3: Training a model on the data ----训练
library(e1071)
sms_classifier <- naiveBayes(sms_train, sms_train_labels)
## Step 4: Evaluating model performance ----预测
sms_test_pred <- predict(sms_classifier, sms_test)
#----评估
library(gmodels)
CrossTable(sms_test_pred, sms_test_labels,
prop.chisq = FALSE, prop.t = FALSE, prop.r = FALSE,
dnn = c('predicted', 'actual'))
## Step 5: Improving model performance ----加入拉普拉斯估计laplace = 1
sms_classifier2 <- naiveBayes(sms_train, sms_train_labels, laplace = 1)
sms_test_pred2 <- predict(sms_classifier2, sms_test)
CrossTable(sms_test_pred2, sms_test_labels,
prop.chisq = FALSE, prop.t = FALSE, prop.r = FALSE,
dnn = c('predicted', 'actual'))
/2 
