R language classification

R Problems encountered in language

• wildcard
  %*% Matrix multiplication

• PCA Principal component analysis

#1 Import data 
data(iris)# Import the built-in data set directly 
head(iris)
#2 Centralize variables （ Subtract the mean value from each data ） And standardization （ And divide it by the standard deviation ）
iris2=scale(iris[,1:4], center=T,scale=T)
head(iris2)
#3 Calculate the covariance matrix 
cm1<-cor(iris2)
cm1
#4 Calculate the eigenvalue matrix , Get eigenvalues and eigenvectors 
rs1<-eigen(cm1)
rs1
eigenvalues <- rs1$values
eigenvector2 <- as.matrix(rs1$vectors)
#5 Calculate the variance contribution of each variable 
(Proportion_of_Variance <- eigenvalues/sum(eigenvalues))
(Cumulative_Proportion <- cumsum(Proportion_of_Variance))
# Drawing the gravel map 
par(mar=c(6,6,2,2))
plot(rs1$values,type="b",
     cex=2,
     cex.lab=2,
     cex.axis=2,
     lty=2,
     lwd=2,
     xlab = "Principal components",
     ylab="Eigenvalues")
# Calculate the principal component score 
dt<-as.matrix(iris2)
PC <- dt %*% eigenvector2
colnames(PC) <- c("PC1","PC2","PC3","PC4")
head(PC)
# Combine principal component scores and category labels 
iris3<-data.frame(PC,iris$V5)
head(iris3)
# Calculate the variance contribution value of the first two principal components 
xlab<-paste0("PC1(",round(Proportion_of_Variance[1]*100,2),"%)")
ylab<-paste0("PC2(",round(Proportion_of_Variance[2]*100,2),"%)")
# Draw the category matrix 
p1<-ggplot(data = iris3,aes(x=PC1,y=PC2,color=iris3[,5]))+
  stat_ellipse(aes(fill=iris3[,5]),
               type ="norm", geom ="polygon",alpha=0.2,color=NA)+
  geom_point()+labs(x=xlab,y=ylab,color="")+
  guides(fill=F)
p1

• LDA discriminant analysis
  And text mining LDA distinguish , Classification of LDA Model refers to projecting data into a discriminant equation , Make the inter class data variance as large as possible , The variance of intra class data should be as small as possible . The maximum number of discriminant equations is min( Number of categories of labels -1, The amount of data that needs to be predicted ), At the same time of classification, dimensionality reduction can also be achieved , The discriminant equation is the new dimension .
  And the previous pca The difference in method is ,pca It is to eliminate variables that have little impact on category labels , And no category labels are required .

#LDA model
f <- paste(names(train_raw.df)[5], "~", paste(names(train_raw.df)[-5], collapse=" + "))# Build regression equation 
iris_raw.lda <- lda(as.formula(paste(f)), data = train_raw.df)
iris_raw.lda.predict <- predict(iris_raw.lda, newdata = test_raw.df)
# Use LDA To make predictions 
pred_y<-iris_raw.lda.predict$class
# draw LDA Prediction chart 
ldaPreds <- iris_raw.lda.predict$x
head(ldaPreds)
test_raw.df %>%
  mutate(LD1 = ldaPreds[, 1],
         LD2 = ldaPreds[, 2]) %>%
  ggplot(aes(LD1, LD2, col = species)) +
  geom_point() +
  stat_ellipse() +
  theme_bw()
# According to the prediction , Calculate the prediction accuracy 
t = table(pred_y,test_y)
acc1 = sum(diag(t))/nrow(test_x) *100
print(paste(" The accuracy of model prediction is ：",round(acc1,4),'%',sep=''))
#ROC
lda_pre2 = predict(iris_raw.lda,test_raw.df,type = "prob")
roc_lda=multiclass.roc(test_y,lda_pre2$posterior)
auc(roc_lda)

• Decision tree
  There are two kinds of regression tree and classification tree , call R In language rpart package . If it is a classification tree, you need to quantify the category label to identify , Subsequent calls predict Function prediction can be selected type yes prob still class, Then we can get the data of posterior probability , Calculation ROC Curve data .

#Decesion Tree
library(rpart)
library(rpart.plot)
library(caret)
train_raw.df$species <- factor(train_raw.df$species)# Category label vectorization 
tree = rpart(species ~ .,data = train_raw.df)# Classification tree model 
summary(tree)
rpart.plot(tree,type = 2)# Draw decision tree 
tree_pre1 = predict(tree,test_raw.df) # Prediction accuracy 
t2 = table(tree_pre1,test_y)
acc2 = sum(diag(t2))/nrow(test_x) *100
print(paste(" The accuracy of model prediction is ：",round(acc2,4),'%',sep=''))
tree_pre2 = predict(tree,test_raw.df,type = "prob")
roc_tree=multiclass.roc(test_y, tree_pre2$posterior)
auc(roc_tree)

• Classification algorithm evaluation index
  Multi classification algorithm calculation ROC（Receiver Operating characteristic Curve） Curve should call pROC Bag multiclass.roc function , Two classification algorithm can be used directly roc function . The multi classification algorithm uses computation roc Generally, you can only get AUC（Multi-class area under the curve） Value , namely ROC The area under the curve enclosed by the coordinate axis ,AUC The value range of is 0.5 and 1 Between .AUC The closer the 1.0, The better the prediction effect of the model ; be equal to 0.5 when , The authenticity is the lowest , No application value . If you want to draw ROC The graph should specify two classification labels .