分类II-神经网络和支持向量机 笔记
支持向量机可以做到全局最优，而神经网络容易陷入多重局部最优。libsvm和SVMLite都是非常流行的支持向量机工具，e1071包提供了libsvm的实现，klap包提供了对后者的实现。
SVM优势在于利用了面向工程问题的核函数，能够提供准确度非常高的模型，同时借助正则项可以避免模型的过度适应，用户不必担心诸如局部最优和多重共线性难题，弊端是训练测试速度慢，模型处理时间冗长，不适合规模庞大数据集。
和神经网络一样，都属于黑盒算法，结果较难解释。另外如何确定合适核函数，也是一个难点，正则化也是需要考虑的问题。gamma函数决定分离超平面的形状，默认为数据维度的倒数，提高它的值通常会增加支持向量的数量。考虑到成本函数，默认值通常为1，此时正则项也是常数，正则项越大，边界越小。
library(e1071)
# install.packages("C50")
# library(C50)
# data('churn', package = 'C50')
# install.packages("modeldata")
# https://stackoverflow.com/questions/60506936/data-set-churn-not-found-in-package-c50
library(modeldata)
data(mlc_churn)
churn <- mlc_churn
churnTrain <- churn[,!names(churn) %in% c('state',
'area_code', "account_length")]
set.seed(2)
ind <- sample(2,nrow(churnTrain),replace = TRUE,
prob = c(0.7,0.3))
trainset <- churnTrain[ind==1,]
testset <- churnTrain[ind==2,]
model <- svm(churn~., data = trainset, kernel = "radial", cost = 1,
gamma=1/ncol(trainset))
summary(model)
Call:
svm(formula = churn ~ ., data = trainset, kernel = "radial",
cost = 1, gamma = 1/ncol(trainset))
# SVMLight
install.packages("klaR")
library(klaR)
unzip("/mnt/c/Users/zd200/OneDrive/Learning/ML/svm_light.zip")
download.file('http://download.joachims.org/svm_light/current/svm_light_linux64.tar.gz',"svm_light_linux64.tar.gz")
untar('svm_light_linux64.tar.gz')
getwd()
model.light <- svmlight(churn~., data = trainset,
kernel="radial", cost=1,gamma=1/ncol(trainset),pathsvm="/mnt/c/Users/zd200/Desktop/Metagenomics/PJS")
'
Parameters:
SVM-Type: C-classification
SVM-Kernel: radial
cost: 1
Number of Support Vectors: 959
( 546 413 )
Number of Classes: 2
Levels:
yes no'
# 选择惩罚因子
# 给iris数据集取子集，用于建模
iris.subset <- subset(iris, select=c("Sepal.Length", "Sepal.Width","Species"),
Species %in% c("setosa","virginica"))
# 绘制散点图
plot(x=iris.subset$Sepal.Length, y=iris.subset$Sepal.Width,
col=iris.subset$Species,pch=1)
# 将惩罚因子设置为1
svm.model <- svm(Species~., data = iris.subset,
kernel="linear", cost=1,scale=FALSE)
# 标注支持向量
points(iris.subset[svm.model$index, c(1,2)], col="blue", cex=2)
# 加分隔线
w <- t(svm.model$coefs) %*% svm.model$SV
b <- -svm.model$rho
abline(a=-b/w[1,2], b=-w[1,1]/w[1,2],col="red", lty=5)
# 惩罚因子1000
# 绘制散点图
plot(x=iris.subset$Sepal.Length, y=iris.subset$Sepal.Width,
col=iris.subset$Species,pch=1)
svm.model <- svm(Species~., data = iris.subset,
kernel="linear", cost=1000,scale=FALSE)
# 标注支持向量
points(iris.subset[svm.model$index, c(1,2)], col="blue", cex=2)
# 加分隔线
w <- t(svm.model$coefs) %*% svm.model$SV
b <- -svm.model$rho
abline(a=-b/w[1,2], b=-w[1,1]/w[1,2],col="red", lty=5)
# SVM可视化
data(iris)
model.iris <- svm(Species ~., iris)
plot(model.iris, iris, Petal.Width~Petal.Length, slice =
list(Sepal.Width = 3, Sepal.Length = 4))
plot(model, trainset, total_day_minutes ~ total_intl_charge)
# 预测分类
svm.pred <- predict(model, testset[,!names(testset) %in% c("churn")])
svm.table <- table(svm.pred, testset$churn)
svm.table
# 调用classAgreement计算分类一致性
classAgreement(svm.table)
# 混淆矩阵
library(caret)
confusionMatrix(svm.table)
两幅图的比较说明，惩罚因子有较大影响。
svm.pred yes no
yes 106 14
no 108 1293
$diag
[1] 0.9197896
$kappa
[1] 0.5936486
$rand
[1] 0.8523496
$crand
[1] 0.5342531
Confusion Matrix and Statistics
svm.pred yes no
yes 106 14
no 108 1293
Accuracy : 0.9198
95% CI : (0.905, 0.9329)
No Information Rate : 0.8593
P-Value [Acc > NIR] : 2.218e-13
Kappa : 0.5936
Mcnemar's Test P-Value : < 2.2e-16
Sensitivity : 0.49533
Specificity : 0.98929
Pos Pred Value : 0.88333
Neg Pred Value : 0.92291
Prevalence : 0.14070
Detection Rate : 0.06969
Detection Prevalence : 0.07890
Balanced Accuracy : 0.74231
'Positive' Class : yes
扩展
# regression
library(car)
library(e1071)
data("Quartet")
model.regression <- svm(Quartet$y1~ Quartet$x, type="eps-regression")
predict.y <- predict(model.regression, Quartet$x)
predict.y
# ############
1 2 3 4 5 6 7 8 9 10 11
8.196894 7.152946 8.807471 7.713099 8.533578 8.774046 6.186349 5.763689 8.726925 6.621373 5.882946
plot(Quartet$x, Quartet$y1,pch=19)
points(Quartet$x, predict.y, pch=15, col='red')
6.6 调整支持向量
除了选择不同的特征集和核函数，还可以借助参数gamma以及惩罚因子来调整支持向量机的性能。tune.svm函数简化了这个过程。
# tuned
tuned <- tune.svm(churn~., data = trainset, gamma = 10^(-6:-1),
cost = 10^(1:2))
summary(tuned)
# ###########
Parameter tuning of ‘svm’:
- sampling method: 10-fold cross validation
- best parameters:
gamma cost
0.01 100
- best performance: 0.07272516
model.tuned <- svm(churn~., data = trainset, gamma =tuned$best.parameters$gamma,
cost = tuned$best.parameters$cost)
summary(model.tuned)
all:
svm(formula = churn ~ ., data = trainset, gamma = tuned$best.parameters$gamma, cost = tuned$best.parameters$cost)
Parameters:
SVM-Type: C-classification
SVM-Kernel: radial
cost: 100
Number of Support Vectors: 768
( 424 344 )
Number of Classes: 2
Levels:
yes no
# predict
svm.tuned.pred <- predict(model.tuned, testset[,!names(testset) %in% c("churn")])
svm.tuned.table <- table(svm.tuned.pred, testset$churn)
svm.tuned.table
#性能评测
classAgreement(svm.tuned.table)
$diag
[1] 0.9250493
$kappa
[1] 0.6504167
$rand
[1] 0.8612426
$crand
[1] 0.5882917
> confusionMatrix(svm.tuned.table)
Confusion Matrix and Statistics
svm.tuned.pred yes no
yes 128 28
no 86 1279
Accuracy : 0.925
95% CI : (0.9106, 0.9378)
No Information Rate : 0.8593
P-Value [Acc > NIR] : 1.027e-15
Kappa : 0.6504
tuned.svm采用十折交叉来获得每次组合的错误偏差，选择误差最低的最佳参数组合。使用这个组合再训练一个支持向量机。
6.7 neuralnet包训练神经网络
我们一般认为神经网络是非常高技术含量的东西，这里我们就学习下这个“高大上”的东西。其实，应该深度学习的技术含量高点，神经网络应该推出好多好多年了。
############################神经网络
data(iris)
ind <- sample(2,nrow(iris),replace = TRUE,
prob = c(0.7,0.3))
trainset <- iris[ind==1,]
testset <- iris[ind==2,]
install.packages("neuralnet")
library(neuralnet)
trainset$setosa <- trainset$Species =='setosa'
trainset$virginica <- trainset$Species=="virginica"
trainset$versicolor <- trainset$Species=='versicolor'
network <- neuralnet(versicolor + virginica + setosa ~ Sepal.Length + Sepal.Width + Petal.Length + Petal.Width, trainset,
hidden = 3)
神经元的优势是可检测非线性关系，利用算法的并行化实现对大数据集的高效训练，无参模型，避免参数估计中的错误。不足是容易陷入局部最优，算法训练时间过长，可能过拟合。
6.8 可视化
# plot
plot(network)
# 可视化泛化权值
# 可视化泛化权值
par(mfrow=c(2,2))
gwplot(network,selected.covariate = "Petal.Width")
gwplot(network,selected.covariate = "Sepal.Width")
gwplot(network,selected.covariate = "Petal.Length")
gwplot(network,selected.covariate = "Sepal.Length")
图中泛化权值接近于0，说明协变量对分类结果影响不大，若总体方差>1，则协变量对分类结果存在非线性影响。
6.9 基于neuralnet包得到的模型实现类标号预测
# 类标号预测
net.predict <- compute(network, testset[-5])$net.result
net.predicttion <- c('setosa', "virginica", 'versicolor')[apply(net.predict,1,which.max)]
predict.table <- table(testset$Species, net.predicttion)
predict.table
net.predicttion
setosa versicolor virginica
setosa 0 14 0
versicolor 8 0 1
virginica 2 0 12
library(e1071)
classAgreement(predict.table)
$diag
[1] 0.3243243
$kappa
[1] -0.004343105
$rand
[1] 0.9099099
$crand
[1] 0.7944529
confusionMatrix(predict.table)
Confusion Matrix and Statistics
net.predicttion
setosa versicolor virginica
setosa 0 14 0
versicolor 8 0 1
virginica 2 0 12
Overall Statistics
Accuracy : 0.3243
95% CI : (0.1801, 0.4979)
No Information Rate : 0.3784
P-Value [Acc > NIR] : 0.8004
Kappa : -0.0043
compute函数还可以获得每一层的输出compute(network, testset[-5])。
6.10 nnet包训练神经模型
这个包提供了传统的前馈反向传播神经网络算法的功能实现，neuralnet包实现了大部分神经网络算法。
# ####nnet
install.packages('nnet')
library(nnet)
# 利用前面分好的训练和测试集 隐藏单元size，初始随机数rang，权值衰减参数decay， 最大迭代次数maxit
iris.nn <- nnet(Species~.,data = trainset[,-c(6,7,8)], size = 2, rang = 0.1, decay = 5e-4, maxit = 200)
summary(iris.nn)
a 4-2-3 network with 19 weights
options were - softmax modelling decay=5e-04
6.11 基于nnet模型实现类预测
library(caret)
iris.predict <- predict(iris.nn, testset[,-c(6,7,8)], type = "class")
m.table <- table(testset$Species, iris.predict)
confusionMatrix(m.table)
confusionMatrix(m.table)
Confusion Matrix and Statistics
iris.predict
setosa versicolor virginica
setosa 14 0 0
versicolor 0 8 1
virginica 0 2 12
Overall Statistics
Accuracy : 0.9189
95% CI : (0.7809, 0.983)
No Information Rate : 0.3784
P-Value [Acc > NIR] : 8.776e-12
Kappa : 0.8768
如果不指定type=class，默认输出概率矩阵。
head(predict(iris.nn, testset[,-c(6,7,8)]))
setosa versicolor virginica
3 0.9995652 0.0004348076 4.156404e-13
4 0.9995187 0.0004813482 4.637527e-13
5 0.9995810 0.0004190092 3.993965e-13
7 0.9995603 0.0004397184 4.206989e-13
17 0.9995906 0.0004094486 3.895892e-13
19 0.9995774 0.0004225847 4.030688e-13
本文参与 腾讯云自媒体分享计划，分享自微信公众号。
原始发表：2021-12-26，如有侵权请联系 cloudcommunity@tencent.com 删除
评论
登录后参与评论
推荐阅读