目录
Intro
近期在整理数据分析流程，找到了之前写的一篇代码，分享给大家。这是我上学时候做的一个项目，当时由于经验不足产生了一些问题，这些问题会在之后一点一点给大家讨论，避免各位踩坑。本篇分享会带一些讲解，可能有些地方不够清楚，欢迎留言讨论。
本次除了分享之外也是对自己之前项目的一个复盘。还是使用R语言（毕竟是我钟爱的语言）。Python的如果有需求之后会放别的项目。
本篇中包含了数据导入，清洗，可视化，特征工程，建模的代码，大家可以选择需要的去参考。
项目背景
数据来自Online Shopper’s Intention 包含12,330 条数据, 10个计数型特征和8个类别型特征。 使用‘Revenue’ 作为标签进行建模。最终目的就是根据拿到的这些数据去建立一个可以预测Revenue的模型。
前期准备
首先你要下载一个R语言以及它的舒适版本R studio。怎么下载呢，把我之前文章上的话直接粘过来哈哈
安装R以及Rstudio
如果之前有用过R的朋友请忽略这一段。
安装R非常简单，直接官网下载
之后下载Rstudio，这个相当于R语言的开挂版，界面相比于R来说非常友好，辅助功能也很多，下载地址
#注意Rstudio是基于R语言的，需要下载安装R语言后才可以安装使用。
安装好了后运行以下代码来导入package们。
setwd("~/Desktop/STAT5003/Ass") #选择项目存放的位置，同样这也是你数据csv存放的位置
# install.packages("xxx") 如果之前没有装过以下的包，先用这句话来装包，然后再去load
# the following packages are for the EDA part
library(GGally)
library(ggcorrplot)
library(psych)
library(ggstatsplot)
library(ggplot2)
library(grid)
# the following packages are for the Model part
library(MASS)
library(Boruta) # Feature selection with the Boruta algorithm
library(caret)
library(MLmetrics)
library(class)
library(neuralnet)
library(e1071)
library(randomForest)
library(keras)
导入的包有些多，keras那个的安装可以参考我之前的文章 （R语言基于Keras的MLP神经网络详解
https://www.jb51.net/article/234031.htm ）
数据描述
首先啊把这个数据下载到你的电脑上，然后用以下代码导入R就可以了。
dataset <- read.csv("online_shoppers_intention.csv")
str(dataset)
str（）这个function可以看到你这个数据的属性，输出如下：
此时发现数据格式有int，number，factor等等。为了之后建分析和建模方便，我们先统一数据格式。
dataset$OperatingSystems <- as.factor(dataset$OperatingSystems) dataset$Browser <- as.factor(dataset$Browser) dataset$Region <- as.factor(dataset$Region) dataset$TrafficType <- as.factor(dataset$TrafficType) dataset$Weekend <- as.factor(dataset$Weekend) dataset$Revenue <- as.factor(dataset$Revenue) dataset$Administrative <- as.numeric(dataset$Administrative) dataset$Informational <- as.numeric(dataset$Informational) dataset$ProductRelated <- as.numeric(dataset$ProductRelated) summary(dataset)
现在数据格式基本统一啦，分为factor和numeric，这方便我们之后的操作。因为R里面的一些package（尤其是建模的package）对数据的输入格式有要求，所以提前处理好非常重要。这可以帮助你更好的整理数据以及敲出简洁舒爽的代码。
记住整理好数据格式之后summary（）一下，你可以从这里发现一些数据的小问题。比如下面的这个‘Administrative_Duration ’。
你看这min=-1就离谱，（当然这也是一个小坑）我们知道duration不可能是<0的。但这是我们的主观思维，由于不知道这个数据在采集入数据库的时候是怎么定义的，所以这个-1是为啥我们不会知道原因。这也是为什么我推荐做数据分析的时候要从头开始跟项目，这样你对数据了如指掌，而不是像现在这样只凭主观思想去判断数据对错（虽然大部分时候你的主观思想没啥问题）
以下给一些数据解释，就不翻译了，看或不看都可（但你自己做项目的时候一定一定一定要仔细看）
Variables are described as follows：
Administrative : Administrative Value
Administrative_Duration : Duration in Administrative Page
Informational : Informational Value
Informational_Duration : Duration in Informational Page
ProductRelated : Product Related Value
ProductRelated_Duration : Duration in Product Related Page
BounceRates : Bounce Rates of a web page
ExitRates : Exit rate of a web page
PageValues : Page values of each web page
SpecialDay : Special days like valentine etc
Month : Month of the year
OperatingSystems : Operating system used
Browser : Browser used
Region : Region of the user
TrafficType : Traffic Type
VisitorType : Types of Visitor
Weekend : Weekend or not
Revenue : Revenue will be generated or not
数据清洗
我们在上一部分的summary已经发现了duration有小于0的，因此所有小于0的duration相关的，我们把它变成NA，然后算一下NA率，来判断这些数是给它填补上还是直接删。个人认为如果missing rate很小删了就成。但如果你的数据集本身就不大，那建议你使用填值法填进去。因为数据太少的话就没啥分析的必要。具体多少算少，见仁见智吧，感兴趣的话之后可以写一篇做讨论。
dataset$Administrative_Duration[dataset$Administrative_Duration < 0] = NA
dataset$Informational_Duration[dataset$Informational_Duration < 0] = NA
dataset$ProductRelated_Duration[dataset$ProductRelated_Duration < 0] = NA
missing.rate <- 1 - nrow(na.omit(dataset))/nrow(dataset)
paste("missing rate =", missing.rate * 100, "%")
"missing rate = 0.381184103811838 %"还挺小的，所以直接删掉有问题的数据。
dataset <- na.omit(dataset)
然后记得用summary再查一次哦，看看是否删干净了。
预分析及预处理
数值型数据
下面三种分别是箱形图，ggpairs以及相关性矩阵。 箱形图可以用来观察数据整体的分布情况。ggpairs绘制的相关关系图可以查看数据分布和相关性。相关性矩阵专注于看相关系数以及是否相关性是否significant。这几个各有其注重点，根据需要去做就可以。
par(mfrow = c(2, 5)) #让图片以2行5列的形式排列在一张图上 boxplot(dataset$Administrative, main = "Administrative") boxplot(dataset$Administrative_Duration, main = "Administrative_Duration") boxplot(dataset$Informational, main = "Informational") boxplot(dataset$Informational_Duration, main = "Informational_Duration") boxplot(dataset$ProductRelated, main = "ProductRelated") boxplot(dataset$ProductRelated_Duration, main = "ProductRelated_Duration") boxplot(dataset$BounceRates, main = "BounceRates") boxplot(dataset$ExitRates, main = "ExitRates") boxplot(dataset$PageValues, main = "PageValues") boxplot(dataset$SpecialDay, main = "SpecialDay")
ggpairs(dataset[, c(1:10)])
corr = cor(dataset[, c(1:10)])
p.mat <- cor_pmat(dataset[, c(1:10)], use = "complete", method = "pearson")
ggcorrplot(corr, hc.order = TRUE, type = "lower", lab = TRUE, p.mat = p.mat,
insig = "blank")
类别型数据
针对类别型数据我们主要是看他的分布，因此直接画bar plot就成。下面的代码用到了ggplot，是个非常好用的可视化包。grid.newpage()这里主要是为了让这些图片都显示在一张图上，这样把图片导出或是直接在markdown上显示的时候所有图都会显示在一个页面上面，看起来比较美观和舒适。
p1 <- ggplot(dataset, aes(x = SpecialDay)) + geom_bar(fill = "#CF6A1A", colour = "black") +
theme_bw()
p2 <- ggplot(dataset, aes(x = Month)) + geom_bar(fill = "#CF6A1A", colour = "black") +
theme_bw()
p3 <- ggplot(dataset, aes(x = OperatingSystems)) + geom_bar(fill = "#CF6A1A",
colour = "black") + theme_bw()
p4 <- ggplot(dataset, aes(x = Browser)) + geom_bar(fill = "#CF6A1A", colour = "black") +
theme_bw()
p5 <- ggplot(dataset, aes(x = Region)) + geom_bar(fill = "#CF6A1A", colour = "black") +
theme_bw()
p6 <- ggplot(dataset, aes(x = TrafficType)) + geom_bar(fill = "#CF6A1A", colour = "black") +
theme_bw()
p7 <- ggplot(dataset, aes(x = VisitorType)) + geom_bar(fill = "#CF6A1A", colour = "black") +
theme_bw()
p8 <- ggplot(dataset, aes(x = Weekend)) + geom_bar(fill = "#CF6A1A", colour = "black") +
theme_bw()
p9 <- ggplot(dataset, aes(x = Revenue)) + geom_bar(fill = "#CF6A1A", colour = "black") +
theme_bw()
grid.newpage()
pushViewport(viewport(layout = grid.layout(4, 3, heights = unit(c(1, 3, 3, 3),
"null"))))
grid.text("Bar Plot of All Categorical Feature", vp = viewport(layout.pos.row = 1,
layout.pos.col = 1:3))
vplayout = function(x, y) viewport(layout.pos.row = x, layout.pos.col = y)
print(p1, vp = vplayout(2, 1))
print(p2, vp = vplayout(2, 2))
print(p3, vp = vplayout(2, 3))
print(p4, vp = vplayout(3, 1))
print(p5, vp = vplayout(3, 2))
print(p6, vp = vplayout(3, 3))
print(p7, vp = vplayout(4, 1))
print(p8, vp = vplayout(4, 2))
print(p9, vp = vplayout(4, 3))
我们可以看到，数据还是比较偏。我们想要预测的revenue也是非常imbalance（标签中的false与true占比不均衡）。因此在处理数据或是选择模型的时候要注意这一点。这里不作详细讨论。针对imbalance data应该是有很多可以说的东西。之后有空的话可以细聊~
其实到目前为止，作为一个普通的项目来说，预分析可以结束了，我们查看了所有数据的分布，并且对现有的数据有了一些直观的印象。但我们不能满足于此，因此对每一个类别型变量再做一次更细致的分析。
首先看一下这个 Special Day 。原数据里给的这个special day给的是0，0.2，0.4这种数值，代表的是距离节日当天的日子，比如1就是节日当天，0.2是节日的前几天（我记得大概是这样）但这种就比较迷惑，我不知道这个具体是咋划分的（这也是为啥希望大家对你所研究的项目有非常深入的了解，你如果对此很了解，那么很多分析的步骤是可以省略的），所以只能让数据告诉我，special day应该如何存在于我们之后的模型中。
special_day_check <- dataset[, c(10, 18)]
special_day_check$Revenue <- ifelse(special_day_check$Revenue == "FALSE", 0,
1)
special_day_check$SpecialDay[special_day_check$SpecialDay == 0] = NA
special_day_check <- na.omit(special_day_check)
special_day_glm <- glm(Revenue ~ SpecialDay, data = special_day_check, family = binomial(link = "logit"))
summary(special_day_glm)
##
## Call:
## glm(formula = Revenue ~ SpecialDay, family = binomial(link = "logit"),
## data = special_day_check)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.3961 -0.3756 -0.3560 -0.3374 2.4491
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -2.3954 0.2986 -8.021 1.05e-15 ***
## SpecialDay -0.5524 0.4764 -1.159 0.246
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 578.11 on 1247 degrees of freedom
## Residual deviance: 576.77 on 1246 degrees of freedom
## AIC: 580.77
##
## Number of Fisher Scoring iterations: 5
首先，我们要检查的是special day 是否应该是一个数值变量。因此，建立一个glm模型（revenue = a+b*special_day)，发现special day的p值=0.246（>0.05），因此可以数值型的认为“SpecialDay”不对revenue有显著的影响，因此specialday可以被当作类别型变量。
现在我们把它当作类别型变量分析一下。用ggbarstats这个function。ggstatsplot是ggplot2包的扩展，主要用于创建美观的图片同时自动输出统计学分析结果，其统计学分析结果包含统计分析的详细信息，该包对于经常需要做统计分析的科研工作者来说非常有用。
ggbarstats(data = dataset, main = Revenue, condition = SpecialDay, sampling.plan = "jointMulti",
title = "Revenue by Special Days", xlab = "Special Days", perc.k = 0.5,
x.axis.orientation = "slant", ggstatsplot.layer = FALSE, messages = FALSE)
用此函数可以绘制出呈现分类变量的柱状图，图中的上半部分( x P e a r s o n 2 x^2_{Pearson} xPearson2, p p p , V C r a m e r V_{Cramer} VCramer 等)代表传统的统计学方法（Frequentist）的一些统计值，下面的部分( l o g e ( B F 01 ) log_e(BF_{01}) loge(BF01)等)代表贝叶斯（Bayesian）的一些统计值。
在本项目中，我们主要关注p-value，我们发现，p<0.001并且在柱状图上方所有都是***，这代表了非常显著。因此我们可以确定special day就这样作为类别型变量使用。
之后把每一个类别型变量都这样做一下。过程不赘述了，挑一个有代表性的给大家看一下。
我们看一下operating systems的ggbarstats（）。
ggbarstats(data = dataset, main = Revenue, condition = OperatingSystems, sampling.plan = "jointMulti",
title = "Revenue by Different Operating Systems", xlab = "Operating Systems",
perc.k = 0.5, x.axis.orientation = "slant", ggstatsplot.layer = FALSE, messages = FALSE)
我们发现整体的p<0.001但是，因为在子类别的样本少，所以柱状图上面出现了ns。我们知道，如果数据很少，那么该数据便不具有统计价值，因此我们把这些少样本的子类别合并在一起，再看一次。
dataset$OperatingSystems <- as.integer(dataset$OperatingSystems)
dataset$OperatingSystems[dataset$OperatingSystems == "5"] <- "other"
dataset$OperatingSystems[dataset$OperatingSystems == "6"] <- "other"
dataset$OperatingSystems[dataset$OperatingSystems == "7"] <- "other"
dataset$OperatingSystems <- as.factor(dataset$OperatingSystems)
ggbarstats(data = dataset, main = Revenue, condition = OperatingSystems, sampling.plan = "jointMulti",
title = "Revenue by Different Operating Systems", xlab = "Operating Systems",
perc.k = 0.5, x.axis.orientation = "slant", ggstatsplot.layer = FALSE, messages = FALSE)
现在看起来就比较舒适了，都很显著。
预处理和预分析到此结束。
特征
我们进行特征工程的最终目的就是提升模型的性能，比如你的数据特征很少的话我们需要建立一些二阶、三阶特征来丰富我们的数据。或是特征太多的时候我们需要进行降维处理。这里我没有做太多的特征工程，只是把特征进行了一下基本的筛选，把没有用的特征删掉。这里的逻辑是先用pca看一下可以保留多少特征，再用Boruta算法和stepAIC去选一下。
# PCA Since pca can only use on numeric data, so we use the os[,c(1:9)] pcdata <- os[, c(1:9)] pclable <- ifelse(os$Revenue == "TRUE", "red", "blue") pc <- princomp(os[, c(1:9)], cor = TRUE, scores = TRUE) summary(pc) ## Importance of components: ## Comp.1 Comp.2 Comp.3 Comp.4 Comp.5 ## Standard deviation 1.8387377 1.2923744 1.0134790 1.0020214 0.9697619 ## Proportion of Variance 0.3756618 0.1855813 0.1141266 0.1115608 0.1044931 ## Cumulative Proportion 0.3756618 0.5612431 0.6753697 0.7869305 0.8914236 ## Comp.6 Comp.7 Comp.8 Comp.9 ## Standard deviation 0.65008195 0.59319914 0.3510795 0.281849096 ## Proportion of Variance 0.04695628 0.03909836 0.0136952 0.008826546 ## Cumulative Proportion 0.93837989 0.97747825 0.9911735 1.000000000 plot(pc, type = "lines")
从pca里面我们可以发现，保留7个numeric变量就可以有95%以上的方差。因此之后我们可以按着至少7个numeric variable这个标准去保留。
Boruta算法
set.seed(123) boruta.train <- Boruta(Revenue ~ ., data = os, doTrace = 2, maxRuns = 15) print(boruta.train) # Boruta performed 14 iterations in 3.920271 mins. 13 attributes confirmed # important: Administrative, Administrative_Duration, BounceRates, Browser, # ExitRates and 8 more; 1 attributes confirmed unimportant: SpecialDay; 2 # tentative attributes left: OperatingSystems, Weekend; so SpecialDay can be # delete when we fit the model. OperatingSystems and Weekend need to check # by other ways.
StepAIC
full.model <- glm(Revenue ~ . - SpecialDay, data = os, family = binomial(link = "logit")) # Backward Stepwise AIC stepback <- stepAIC(full.model, direction = "backward", steps = 3) summary(stepback) # OperatingSystems, Weekend are all above the <none>, combine the previous # result by Boruta algorithm, it can be delete when we fit model. Browser # has the minimum AIC, it can be delete when we fit model. PCA shows we # should keep 7 numeric variables in the dataset when fit the model, so two # numeric variables should be remove. Informational_Duration and # Administrative has the minimum AIC in numeric variables, so remove these # two variables.
综合上面三个特征选择的方法 SpecialDay, OperatingSystems, Weekend, Browser, Informational_Duration 和 Administrative 应当在建模的时候被移除。有兴趣的可以跑一下上面的代码，由于运行时间有点长，结果就直接码在代码框里了。
建模
现在把用来建模数据整理好，准备建模。
os_modeldata <- os[, -c(1, 4, 10, 11, 12, 16)] # summary(os_modeldata) write.csv(os_modeldata, "os_modeldata.csv")
首先划分训练集和测试集（train 和 test）
set.seed(123)
os_modeldata <- read.csv("os_modeldata.csv")
os_modeldata <- os_modeldata[, -1]
os_modeldata$Revenue <- as.factor(os_modeldata$Revenue)
inTrain <- createDataPartition(os_modeldata$Revenue, p = 0.9)[[1]]
Train <- os_modeldata[inTrain, ]
Test <- os_modeldata[-inTrain, ]
然后把训练集拆成train和val。这里加了个10-cv。有些模型的function可以自己加cv，但由于要用到不同的建模package，为了避免不同package之间划分cv的差异，咱自己建~
add_cv_cohorts <- function(dat, cv_K) {
if (nrow(dat)%%cv_K == 0) {
# if perfectly divisible
dat$cv_cohort <- sample(rep(1:cv_K, each = (nrow(dat)%/%cv_K)))
} else {
# if not perfectly divisible
dat$cv_cohort <- sample(c(rep(1:(nrow(dat)%%cv_K), each = (nrow(dat)%/%cv_K +
1)), rep((nrow(dat)%%cv_K + 1):cv_K, each = (nrow(dat)%/%cv_K))))
}
return(dat)
}
# add 10-fold CV labels to real estate data
train_cv <- add_cv_cohorts(Train, 10)
# str(train_cv)
首先建一个基准模型，Logistic regression classifer（benchmark model）
train_cv_glm <- train_cv
glm.acc <- glm.f1 <- c()
train_cv_glm$Revenue <- ifelse(train_cv_glm$Revenue == "TRUE", 1, 0)
# str(train_cv_glm)
for (i in 1:10) {
# Segement my data by fold using the which() function
indexes <- which(train_cv_glm$cv_cohort == i)
train <- train_cv_glm[-indexes, ]
val <- train_cv_glm[indexes, ]
# Model
glm.model <- glm(Revenue ~ . - cv_cohort, data = train, family = binomial(link = "logit"))
# predict
glm.pred <- predict(glm.model, newdata = val, type = "response")
glm.pred <- ifelse(glm.pred > 0.5, 1, 0)
# evaluate
glm.f1[i] <- F1_Score(val$Revenue, glm.pred, positive = "1")
glm.acc[i] <- sum(glm.pred == val$Revenue)/nrow(val)
}
# F1 and ACC
glm.acc.train <- round(mean(glm.acc), 5) * 100
glm.f1.train <- round(mean(glm.f1), 5) * 100
# print(glm.cm <- table(glm.pred, val$Revenue))
paste("The accuracy by Logistic regression classifier by 10-fold CV in train data is",
glm.acc.train, "%")
paste("The F1-score by Logistic regression classifier by 10-fold CV in train data is",
glm.f1.train, "%")
# f1 = 0.50331
然后建立我们用来对比的机器学习模型。这里使用网格搜索法调参。
KNN
# since knn() function can't use factor as indenpent variable So re-coding
# data, factor to dummy variable)
train_cv_knn <- as.data.frame(model.matrix(~., train_cv[, -11]))
train_cv_knn$Revenue <- train_cv$Revenue
train_cv_knn <- train_cv_knn[, -1]
# head(train_cv_knn)
knn.grid <- expand.grid(k = c(1:30))
knn.grid$acc <- knn.grid$f1 <- NA
knn.f1 <- knn.acc <- c()
for (k in 1:nrow(knn.grid)) {
for (i in 1:10) {
# Segement my data by fold using the which() function
indexes <- which(train_cv_knn$cv_cohort == i)
train <- train_cv_knn[-indexes, ]
val <- train_cv_knn[indexes, ]
# model and predict
knn.pred <- knn(train[, -c(34, 35)], val[, -c(34, 35)], train$Revenue,
k = k)
# evaluate
knn.f1[i] <- F1_Score(val$Revenue, knn.pred, positive = "TRUE")
knn.acc[i] <- sum(knn.pred == val$Revenue)/nrow(val)
}
knn.grid$f1[k] <- mean(knn.f1)
knn.grid$acc[k] <- mean(knn.acc)
print(paste("finished with =", k))
}
print(knn.cm <- table(knn.pred, val$Revenue))
knn.grid[which.max(knn.grid$f1), ]
# k = 7, f1=0.5484112, acc=0.885042
SVM
svm.grid <- expand.grid(cost = c(0.1, 1, 10), gamma = seq(0.2, 1, 0.2))
svm.grid$acc <- svm.grid$f1 <- NA
svm.f1 <- svm.acc <- c()
for (k in 1:nrow(svm.grid)) {
for (i in 1:10) {
# Segement my data by fold using the which() function
indexes <- which(train_cv$cv_cohort == i)
train <- train_cv[-indexes, ]
val <- train_cv[indexes, ]
# model
svm.model <- svm(Revenue ~ ., kernel = "radial", type = "C-classification",
gamma = svm.grid$gamma[k], cost = svm.grid$cost[k], data = train[,
-12])
svm.pred <- predict(svm.model, val[, -12])
# evaluate
svm.f1[i] <- F1_Score(val$Revenue, svm.pred, positive = "TRUE")
svm.acc[i] <- sum(svm.pred == val$Revenue)/nrow(val)
}
svm.grid$f1[k] <- mean(svm.f1)
svm.grid$acc[k] <- mean(svm.acc)
print(paste("finished with:", k))
}
print(svm.cm <- table(svm.pred, val$Revenue))
svm.grid[which.max(svm.grid$f1), ]
# cost=1, gamma=0.2,f1= 0.5900601,acc= 0.8948096
Random Forest
rf.grid <- expand.grid(nt = seq(100, 500, by = 100), mrty = c(1, 3, 5, 7, 10))
rf.grid$acc <- rf.grid$f1 <- NA
rf.f1 <- rf.acc <- c()
for (k in 1:nrow(rf.grid)) {
for (i in 1:10) {
# Segement my data by fold using the which() function
indexes <- which(train_cv$cv_cohort == i)
train <- train_cv[-indexes, ]
val <- train_cv[indexes, ]
# model
rf.model <- randomForest(Revenue ~ ., data = train[, -12], n.trees = rf.grid$nt[k],
mtry = rf.grid$mrty[k])
rf.pred <- predict(rf.model, val[, -12])
# evaluate
rf.f1[i] <- F1_Score(val$Revenue, rf.pred, positive = "TRUE")
rf.acc[i] <- sum(rf.pred == val$Revenue)/nrow(val)
}
rf.grid$f1[k] <- mean(rf.f1)
rf.grid$acc[k] <- mean(rf.acc)
print(paste("finished with:", k))
}
print(rf.cm <- table(rf.pred, val$Revenue))
rf.grid[which.max(rf.grid$f1), ]
# nt=200,mtry=3 ,f1 = 0.6330392, acc=0.8960723
Neural Network
nndata <- Train
nndata$Revenue <- ifelse(nndata$Revenue == "TRUE", 1, 0)
train_x <- model.matrix(~., nndata[, -11])
train_x <- train_x[, -1]
train_y <- to_categorical(as.integer(as.matrix(array(nndata[, 11]))), 2)
model <- keras_model_sequential()
# defining model's layers
model %>% layer_dense(units = 30, input_shape = 33, activation = "relu") %>%
layer_dense(units = 40, activation = "relu") %>% layer_dropout(rate = 0.4) %>%
layer_dense(units = 60, activation = "relu") %>% layer_dropout(rate = 0.4) %>%
layer_dense(units = 30, activation = "relu") %>% layer_dropout(rate = 0.4) %>%
layer_dense(units = 2, activation = "sigmoid")
# defining model's optimizer
model %>% compile(loss = "binary_crossentropy", optimizer = "adam", metrics = c("accuracy"))
# Metrics: The performance evaluation module provides a series of functions
# for model performance evaluation. We use it to determine when the NN
# should stop train. The ultimate measure of performance is F1.
# Check which column in train_y is FALSE
table(train_y[, 1]) # the first column is FALSE
table(train_y[, 1])[[2]]/table(train_y[, 1])[[1]]
# Define a dictionary with your labels and their associated weights
weight = list(5.5, 1) # the proportion of FALSE and TURE is about 5.5:1
# fitting the model on the training dataset
model %>% fit(train_x, train_y, epochs = 50, validation_split = 0.2, batch_size = 512,
class_weight = weight)
# after epoch = 20, val_loss not descrease and val_acc not increase, so NN
# should stop at epoch = 20
模型对比
GLM
glmdata <- Train
glmdata$Revenue <- ifelse(glmdata$Revenue == "TRUE", 1, 0)
testglm <- Test
testglm$Revenue <- ifelse(testglm$Revenue == "TRUE", 1, 0)
glm.model.f <- glm(Revenue ~ ., data = glmdata, family = binomial(link = "logit"))
glm.pred.f <- predict(glm.model.f, newdata = Test, type = "response")
glm.pred.f <- ifelse(glm.pred.f > 0.5, 1, 0)
glm.f1.f <- F1_Score(testglm$Revenue, glm.pred.f, positive = "1")
paste("The F1-score by Logistic regression classifier in test data is", glm.f1.f)
KNN
knndata <- as.data.frame(model.matrix(~., Train[, -11]))
knndata <- knndata[, -1]
knntest <- as.data.frame(model.matrix(~., Test[, -11]))
knntest <- knntest[, -1]
knn.model.f.pred <- knn(knndata, knntest, Train$Revenue, k = 7)
knn.f1.f <- F1_Score(Test$Revenue, knn.model.f.pred, positive = "TRUE")
paste("The F1-score by KNN classifier in test data is", knn.f1.f)
SVM
svm.model.f <- svm(Revenue ~ ., kernel = "radial", type = "C-classification",
gamma = 0.2, cost = 1, data = Train)
svm.pred.f <- predict(svm.model.f, Test)
svm.f1.f <- F1_Score(Test$Revenue, svm.pred.f, positive = "TRUE")
paste("The F1-score by SVM classifier in test data is", svm.f1.f)
Random Forests
rf.model.f <- randomForest(Revenue ~ ., data = Train, n.trees = 200, mtry = 3)
rf.pred.f <- predict(rf.model.f, Test)
rf.f1.f <- F1_Score(Test$Revenue, rf.pred.f, positive = "TRUE")
paste("The F1-score by Random Forests classifier in test data is", rf.f1.f)
NN
nndata <- Train
nndata$Revenue <- ifelse(nndata$Revenue == "TRUE", 1, 0)
train_x <- model.matrix(~., nndata[, -11])
train_x <- train_x[, -1]
train_y <- to_categorical(as.integer(as.matrix(array(nndata[, 11]))), 2)
model <- keras_model_sequential()
# defining model's layers
model %>% layer_dense(units = 30, input_shape = 33, activation = "relu") %>%
layer_dense(units = 40, activation = "relu") %>% layer_dropout(rate = 0.4) %>%
layer_dense(units = 60, activation = "relu") %>% layer_dropout(rate = 0.4) %>%
layer_dense(units = 30, activation = "relu") %>% layer_dropout(rate = 0.4) %>%
layer_dense(units = 2, activation = "sigmoid")
# defining model's optimizer
model %>% compile(loss = "binary_crossentropy", optimizer = "adam", metrics = c("accuracy"))
weight = list(5.5, 1)
model %>% fit(train_x, train_y, epochs = 20, batch_size = 512, class_weight = weight)
# test data
testnn <- Test
testnn$Revenue <- ifelse(testnn$Revenue == "TRUE", 1, 0)
test_x <- model.matrix(~., testnn[, -11])
test_x <- test_x[, -1]
nn.pred <- model %>% predict(test_x)
nn.pred <- as.data.frame(nn.pred)
nn.pred$label <- NA
nn.pred$label <- ifelse(nn.pred$V2 > nn.pred$V1, "TRUE", "FALSE")
nn.pred$label <- as.factor(nn.pred$label)
nn.f1 <- F1_Score(Test$Revenue, nn.pred$label, positive = "TRUE")
paste("The F1-score by Neural network in test data is", nn.f1)
看一下结果对比哈，RF和NN的表现较好。最后做个混淆矩阵看一下。
# RF print(rf.cm.f <- table(rf.pred.f, Test$Revenue)) ## ## rf.pred.f FALSE TRUE ## FALSE 987 74 ## TRUE 50 116 # NN print(nn.cm.f <- table(nn.pred$label, Test$Revenue)) ## ## FALSE TRUE ## FALSE 980 69 ## TRUE 57 121
