1. 技术文章
  2. Plotting a linear discriminant analysis, classification tree and Naive Bayes Curve on a single ROC plot

Plotting a linear discriminant analysis, classification tree and Naive Bayes Curve on a single ROC plot

由 为君一笑 提交于 2019-12-06 22:17:51

First of all, one of the most important issues regarding subsetting your data into training and testing subsets is prior to subsetting, the data have to be randomized otherwise you will have unequal division of your categories in the training and testing data subsets.

Some notes on the code below. I have used the caret package for simplicity for the model fitting methods.

library(pROC)
library(MASS)
library(caret)

set.seed(1234)

mydat <- read.table("~/Desktop/family.txt", header = TRUE, stringsAsFactors= FALSE)
mydat$Family <- factor(mydat$Family, levels = c("v4", "G8"))

# Randomly permute the data before subsetting
mydat_idx <- sample(1:nrow(mydat), replace = FALSE)
mydat <- mydat[mydat_idx, ]

mydat_resampled_idx <- createDataPartition(mydat_idx, times = 1, p = 0.7, list = FALSE)
mydat_resampled <- mydat[mydat_resampled_idx, ] # Training portion of the data

Linear discriminant analysis

lda_mod <-train(x = mydat_resampled[, 2:9], y = as.factor(mydat_resampled[, 1]),
  method = "lda", trControl = trainControl(method = "cv", classProbs = TRUE))

# Generate model predictions 
lda_pred <- predict(lda_mod, newdata = mydat[ , 2:9], type = "prob")

# Store the predictions with the data set
mydat['lda_pred'] <- lda_pred["G8"] # Here we only want the probability associated
                                    # with the class (Y = 1), or in this case, G8

Naive Bayes

nb_tune <- data.frame(usekernel =TRUE, fL = 0)
nb_mod <- train(x = mydat_resampled[, 2:9], y = as.factor(mydat_resampled[, 1]), 
    method = "nb", trControl = trainControl(method = "cv", classProbs = TRUE), tuneGrid = nb_tune)

# Model predictions
nb_pred <- predict(nb_mod, newdata = mydat[ , 2:9], type = "prob")
mydat['nb_pred'] <- nb_pred["G8"]

Classification tree

ct_mod <- train(x = mydat_resampled[, 2:9], y = as.factor(mydat_resampled[, 1]), 
method = "rpart", trControl = trainControl(method = "cv", classProbs = TRUE))
ct_pred <- predict(ct_mod, newdata = mydat[ , 2:9], type = "prob")
mydat['ct_pred'] <- ct_pred["G8"]

ROC curves on the training and testing portions of the data

EDIT: Changed the calculation and plots of AUC curves to use the pROC package 

mydat$binary_response <- as.numeric(mydat$Family) - 1 # convert factor to 0, 1 with G8  = 1

lda_train_roc <- roc(binary_response ~ lda_pred, data = mydat[mydat_resampled_idx, ], ci = TRUE)
nb_train_roc <- roc(binary_response ~ nb_pred, data =  mydat[mydat_resampled_idx, ], ci = TRUE)
ct_train_roc <- roc(binary_response ~ ct_pred, data =  mydat[mydat_resampled_idx, ], ci = TRUE)

lda_test_roc <- roc(binary_response ~ lda_pred, data =  mydat[-mydat_resampled_idx, ], ci = TRUE)
nb_test_roc <- roc(binary_response ~ nb_pred, data =  mydat[-mydat_resampled_idx, ], ci = TRUE)
ct_test_roc <- roc(binary_response ~ ct_pred, data =  mydat[-mydat_resampled_idx, ], ci = TRUE)


par(mfrow = c(2, 1))
plot(lda_train_roc, las = 1, main = "Training data")
plot(nb_train_roc, add = TRUE, col = "red")
plot(ct_train_roc, add = TRUE, col = "blue")
legend(0.4, 0.4, legend = c("lda", "nb", "ct"), lty = c(1,1,1), col = c("black", "red", "blue"))

plot(lda_test_roc, las = 1, main = "Testing data")
plot(nb_test_roc, add = TRUE, col = "red")
plot(ct_test_roc, add = TRUE, col = "blue")
legend(0.4, 0.4, legend = c("lda", "nb", "ct"), lty = c(1,1,1), col =     c("black", "red", "blue"))
par(mfrow = c(1, 1))

# AUC with 95% CL
lda_train_roc$ci[c(2, 1, 3)] #  0.8353741 0.7235472 0.9472011

nb_train_roc$ci[c(2, 1, 3)] # 0.9714286 [0.9303684, 1.0000000]

ct_train_roc$ci[c(2, 1, 3)] # 0.7619048 [0.6524637, 0.8713458]
lda_test_roc$ci[c(2, 1, 3)] # 0.6148148 [0.3555396, 0.8740900]

nb_test_roc$ci[c(2, 1, 3)] # 0.7407407 [0.5345984, 0.9468831]
ct_test_roc$ci[c(2, 1, 3)] # 0.6000000 [0.4139795, 0.7860205]

标签
r
classification
roc
naivebayes
auc
易学教程内所有资源均来自网络或用户发布的内容，如有违反法律规定的内容欢迎反馈！
该文章没有解决你所遇到的问题?点击提问,说说你的问题,让更多的人一起探讨吧!