Machine Learning for hackers读书笔记(六)正则化：文本回归

data<-'F:\\learning\\ML_for_Hackers\\ML_for_Hackers-master\\06-Regularization\\data\\'

ranks <- read.csv(file.path(data, 'oreilly.csv'),stringsAsFactors = FALSE)

library('tm')

documents <- data.frame(Text = ranks$Long.Desc.)
row.names(documents) <- 1:nrow(documents)

#获得语料库

corpus <- Corpus(DataframeSource(documents))

#R2版本用corpus <- tm_map(corpus, tolower)

corpus <- tm_map(corpus, content_transformer(tolower))

#R2版本用corpus <- tm_map(corpus, stripWhitespace)

corpus <- tm_map(corpus, content_transformer(stripWhitespace))

#去除英文停用词
corpus <- tm_map(corpus, removeWords, stopwords('english'))

#得到词项文档矩阵

dtm <- DocumentTermMatrix(corpus)

x <- as.matrix(dtm)
y <- rev(1:100)  #反转1..100，结果是100..1

set.seed(1)

library('glmnet')

performance <- data.frame()

for (lambda in c(0.1, 0.25, 0.5, 1, 2, 5))
{
for (i in 1:50)
{
indices <- sample(1:100, 80)

training.x <- x[indices, ]
training.y <- y[indices]

test.x <- x[-indices, ]
test.y <- y[-indices]

glm.fit <- glmnet(training.x, training.y)

predicted.y <- predict(glm.fit, test.x, s = lambda)

rmse <- sqrt(mean((predicted.y - test.y) ^ 2))

performance <- rbind(performance,data.frame(Lambda = lambda,Iteration = i,RMSE = rmse))
}
}

ggplot(performance, aes(x = Lambda, y = RMSE)) +stat_summary(fun.data = 'mean_cl_boot', geom = 'errorbar') +

stat_summary(fun.data = 'mean_cl_boot', geom = 'point')

#从图上看，失败

#失败了作分类，判断一本书能不能进前50

y <- rep(c(1, 0), each = 50)

#作逻辑回归

regularized.fit <- glmnet(x, y, family = 'binomial')

#预测一下

predict(regularized.fit, newx = x, s = 0.001)

#出来的结果并不是分类，而是一堆数值，因此改一下

ifelse(predict(regularized.fit, newx = x, s = 0.001) > 0, 1, 0)

#第二种方法，把预测结果转成概率值

library('boot')

inv.logit(predict(regularized.fit, newx = x, s = 0.001))

#看效果

set.seed(1)

performance <- data.frame()

for (i in 1:250)
{
indices <- sample(1:100, 80)

training.x <- x[indices, ]
training.y <- y[indices]

test.x <- x[-indices, ]
test.y <- y[-indices]

for (lambda in c(0.0001, 0.001, 0.0025, 0.005, 0.01, 0.025, 0.5, 0.1))
{
glm.fit <- glmnet(training.x, training.y, family = 'binomial')
predicted.y <- ifelse(predict(glm.fit, test.x, s = lambda) > 0, 1, 0)
error.rate <- mean(predicted.y != test.y)

performance <- rbind(performance,data.frame(Lambda = lambda,Iteration = i,ErrorRate = error.rate))

}
}

#画个图
ggplot(performance, aes(x = Lambda, y = ErrorRate)) +
stat_summary(fun.data = 'mean_cl_boot', geom = 'errorbar') +
stat_summary(fun.data = 'mean_cl_boot', geom = 'point') +scale_x_log10()