#加载R包library(corrplot)library(VIM)library(ggplot2)library(gridExtra)library(scales)library(tidyr)library(dplyr)library(readr)library(ggformula)library(imputeMissings)library(gapminder)library(caret)library(rpart)library(rattle)library(kernlab)library(randomForest)library(DMwR2)library(car)library(MLeval)#导入数据#然后就是加载上一次分析得到的处理后的数据clinical <- read.csv("clinical.csv", header=TRUE, sep=",", stringsAsFactors = FALSE)clinical_e<- clinicaldim(clinical_e)

#然后我们需要重新定义一下木匾变量，就是overall_survival列make.names("overall_survival", unique = TRUE, allow_ = FALSE); clinical_e  %>%   mutate(overall_survival = factor(overall_survival,           labels = make.names(levels(overall_survival))));levels(clinical_e$overall_survival) <- c("D", "L")#死亡就是D，存活就是L#本次分析呢，就是通过随机森林和支持向量机 (SVM) 预测癌症患者是否会生存，其中SVM包含两种核函数即多项式核和径向核。#首先我们要在随机森林和SVM上进行单层10倍交叉验证。选择每次分割 (mtry) 时随机采样作为候选变量的数量来优化随机森林模型。“degree of the polynomial”和cost分别在多项式和径向核中分别进行调整。#话不多说，创建模型吧！set.seed(789)#设置随机种子#定义交叉验证参数ctrl = trainControl(method = "cv", number = 10,                     savePredictions = TRUE,                     classProbs = TRUE,                     verboseIter = TRUE)#定义一个非分类的交叉验证参数ctrl_1 = trainControl(method = "cv", number = 10)#保留一下数据data_used = clinical_edata_used_1 = clinical#设置多项式核SVM交叉验证fit_poly = train(overall_survival ~ .,               data = data_used,               na.action = na.exclude,               method = "svmPoly",tuneGrid = expand.grid(C = c(1), degree = seq(0, 5, by = 1), scale = c(1)), # gammapreProcess = c("center","scale"),trControl = ctrl)#看一下交叉验证结果，给出了最优参数，就在图片最后一行fit_poly

#设置径向核SVM交叉验证fit_radial = train(overall_survival ~ .,data = data_used,method = "svmRadial",tuneGrid = expand.grid(C = c(1:5), sigma = c(1)),preProcess = c("center","scale"),trControl = ctrl)fit_radial

#设置随机森林交叉验证fit_rf = train(overall_survival ~.,data = data_used,method = "rf",na.action = na.exclude,tuneGrid = expand.grid(mtry = c(3:12)),preProcess = c("center","scale"),trControl = ctrl)fit_rf

#定义不分类随机森林交叉验证，这个结果没有上面那个好，那就选上面那个fit_rf_1 = train(overall_survival ~.,data = data_used_1,method = "rf",na.action = na.exclude,tuneGrid = expand.grid(mtry = c(3:12)),preProcess = c("center","scale"),trControl = ctrl_1)

all_best_Types = c("Polynomial Kernel","Radial Kernel","Random forests")  all_best_Pars = list(fit_poly$bestTune, fit_radial$bestTune,fit_rf$bestTune )  all_best_Models = list(fit_poly,fit_radial,fit_rf)  all_best_Accuracy = c(max(fit_poly$results$Accuracy),max(fit_radial$results$Accuracy),                        max(fit_rf$results$Accuracy))    ############# 模型对比#############  # 看看特征数和准确率的关系  mtree = length(seq(0, 5, by = 1));   mtree1 = length(seq(1, 5, by = 1))  mtree2 = length(seq(3, 12, by = 1))  mmodels = mtree+mtree1+mtree2  modelMethod = c(rep("Polynomial Kernel",mtree),rep("Radial Kernel",mtree1), rep("Random forests",mtree2))  all_caret_Accuracy = c((fit_poly$results$Accuracy),fit_radial$results$Accuracy, fit_rf$results$Accuracy)  coloptions = rainbow(4)  colused = coloptions[as.numeric(factor(modelMethod))]  charused = 5*(as.numeric(factor(modelMethod)))#可视化结果看看plot(1:mmodels,all_caret_Accuracy,col=colused,pch=charused,     xlab = "Model label",ylab = "Accuracy")  order.min = c(which.max(fit_poly$results$Accuracy),mtree+which.max(fit_radial$results$Accuracy),mtree+mtree                -1+which.max(fit_rf$results$Accuracy))  abline(v=order.min,lwd=0.5)  abline(v=order.min[2],col="red",lwd=0.5)  abline(v=order.min[3],col="green",lwd=0.5)

#整体看的话，还是随机森林要更好一点，多项式核SVM不太稳定，径向核SVM稳定，但是ACC太低了#然后就要画一下ROC曲线了，毕竟这是大家最常用的评估手段了#这里会产生很多图，可以获取每个模型的关键信息，这里大海哥就不展示了poly <- evalm(fit_poly);radial <- evalm(fit_radial);rf <- evalm(fit_rf);#获取我们想要的信息fpr = poly$roc$data$FPRpo = poly$roc$data$SENSra = radial$roc$data$SENSrff = rf$roc$data$SENSroc_fig <- data_frame(fpr,po,ra, rff)
require(ggplot2)colors <- c("Polynomial Kernel" = "blue", "Radial Kernel" = "red", "Random forests" = "green")ggplot(roc_fig, aes(fpr)) +                      geom_line(aes(y=po, colour="Polynomial Kernel"), size = 1) +    geom_line(aes(y=ra, colour="Radial Kernel"), size = 1)+    geom_line(aes(y=rff, colour="Random forests"), size = 1)+    geom_abline(intercept = 0, slope = 1, col = "gray")+  labs(title ="ROC curve",    x = "Specificities",         y = "Sensitivities",         color = "Legend") +    scale_color_manual(values = colors)

#这里看ROC，多项式核SVM和随机森林不相伯仲！Model = c("Polynomial Kernel","Radial Kernel","Random forests")AUC_Val  = c(poly$stdres$`Group 1`$Score[13],radial$stdres$`Group 1`$Score[13], rf$stdres$`Group 1`$Score[13])#同时我们还整合了ACC值，一起看一下df2 <- data.frame(supp=rep(c("Accuracy", "AUC"), each=3),                mod=rep(Model,2),                va=c(all_best_Accuracy, AUC_Val))ggplot(data=df2, aes(x=mod, y=va, fill=supp)) +geom_bar(stat="identity", position=position_dodge())+  labs(x = "",         y = "Value",fill = "")

#但是ACC好像随机森林更优！#看看随机森林的重要性评分importance(fit_rf_1$finalModel)varImpPlot(fit_rf$finalModel)