陈强教授《机器学习及R应用》课程作业

#第十五章作业 #15.1预测Ablone的rings数据 #（1）把因子变量变成虚拟变量 rm(list=ls()) setwd("/Users/Antony/OneDrive/CoreTech/ML&R/MLR_Data") library(AppliedPredictiveModeling) data("abalone") model.matrix(ablone$Tpye,data = abalone) summary(abalone) dummy <- model.matrix(~Type,data=abalone) aba_data <- cbind(abalone[,-1],dummy[,-1]) y.test <- abalone[-train_index,"Rings"] #（2）数据预处理，所有变量进行归一化处理 maxs <- apply(aba_data,2,max) mins <- apply(aba_data, 2, min) aba_s <- as.data.frame(scale(aba_data,center = mins,scale=maxs-mins)) #（3）选取1000个变量作为训练集 set.seed(1) train_index <- sample(4177,1000) train <- aba_s[train_index,] test <- aba_s[-train_index,] fit <- neuralnet(Rings~.,data = train,hidden = 3,act.fct = "logistic") #（4）计算均方误差 plot(fit,fontsize = 10) pred <- predict(fit,test) pred <- pred*(max(abalone$Rings)-min(abalone$Rings))+min(abalone$Rings) mean((pred-y.test)^2) #（5）线性回归测试误差对比 fit.ols <- lm(Rings~., data=train) pred.ols <- predict(fit.ols,test) pred.ols <- pred.ols*(max(abalone$Rings)-min(abalone$Rings))+min(abalone$Rings) mean((pred.ols-y.test)^2) par(mfrow=c(1,2)) plot(pred,y.test,xlab="Prediction",ylab="rings",main="Neural Network") abline(0,1) plot(pred.ols,y.test,xlab="Prediction",ylab="rings",main="OLS Regression") abline(0,1) par(mfrow=c(1,1)) #（6）测试最小误差的神经元数目 MSE <- numeric(10) for (i in 1:10){ set.seed(123) fit <- neuralnet(Rings~.,data=train,hidden=i,act.fct = "logistic") pred <- predict(fit,test) pred <- pred*(max(abalone$Rings)-min(abalone$Rings))+min(abalone$Rings) MSE[i] <- mean((pred-y.test)^2) } min(MSE) which.min(MSE) plot(1:10,MSE,type = "b",xlab="hidden layers") #15.2 使用R包mlbench中Ionosphere（电离层）数据，估计二分类问题 #（1）数据预处理：class变成因子变量，去掉无变化的V2，把两个水平的因子变量V1变成虚拟变量 rm(list=ls()) library(mlbench) data("Ionosphere") Ionosphere <- subset(Ionosphere,select = -c(V2)) #变量改名 library(reshape) names(Ionosphere)[1]="V1-" #利用package处理哑变量 library(caret) str(Ionosphere) dmi <- dummyVars(~.,data=Ionosphere) dmio <- data.frame(predict(dmi,newdata=Ionosphere)) #（2）选取训练集，进行训练 set.seed(1) train <- sample(351,251) fit <- neuralnet(Class.good~.,data=dmio[train,],hidden=5,act.fct = "logistic") #（3）测试集均方误差 prob <- predict(fit,dmio[-train,]) pred <- prob>0.5 (table <- table(Predicted=pred,Actual=dmio[-train,"Class.good"])) (error <- 1-sum(diag(table))/sum(table)) #（4）对比逻辑回归 library(nnet) fit.l <- multinom(Class.good~.,data=dmio[train,]) pred <- predict(fit.l,type="class",newdata = dmio[-train,]) (table <- table(Predicted=pred,Actual=dmio[-train,"Class.good"])) (error <- 1-sum(diag(table))/sum(table)) #（5）在1:10之间for循环 MSE <- numeric(10) for (i in 1:10){ set.seed(123) fit <- neuralnet(Class.good~.,data=dmio[train,],hidden=i,act.fct = "logistic") prob <- predict(fit,dmio[-train,]) pred <- prob>0.5 (table <- table(Predicted=pred,Actual=dmio[-train,"Class.good"])) MSE[i] <- 1-sum(diag(table))/sum(table) } min(MSE) which.min(MSE) plot(1:10,MSE,type = "b",xlab="hidden layers")