如何用R的glmnet包获取混淆矩阵、概率图并设置最优截断值？

1. 生成训练集与测试集的预测概率 #

基于最优lambda值（lambda.min）生成预测概率，注意测试集自变量需与训练集保持一致格式（用model.matrix处理）：

# 生成训练集预测概率
train_pred_prob <- predict(cvfit, newx = x, s = "lambda.min", type = "response")

# 处理测试集自变量
x_test <- model.matrix(asthma ~ temperature+bmi, data = test)[,-1]
# 生成测试集预测概率
test_pred_prob <- predict(cvfit, newx = x_test, s = "lambda.min", type = "response")

2. 绘制预测概率分布图 #

用ggplot2绘制不同真实类别下的预测概率分布，直观展示模型区分能力：

library(ggplot2)

# 整理训练集绘图数据
train_plot_data <- data.frame(
  true_label = factor(train$asthma, levels = c(0,1), labels = c("非哮喘", "哮喘")),
  pred_prob = as.vector(train_pred_prob)
)

# 绘制密度图
ggplot(train_plot_data, aes(x = pred_prob, fill = true_label)) +
  geom_density(alpha = 0.5) +
  labs(x = "预测概率", y = "密度", title = "训练集预测概率分布", fill = "真实类别") +
  theme_minimal()

3. 寻找同时最大化敏感性和特异性的截断值 #

通过**约登指数（Youden's J statistic）**确定最优截断值，约登指数为敏感性 + 特异性 - 1，取该指数最大值对应的阈值即可，使用pROC包实现：

library(pROC)

# 构建测试集ROC曲线对象
roc_obj <- roc(test$asthma, test_pred_prob)

# 计算约登指数最大的截断值
optimal_thresh <- coords(roc_obj, "best", ret = "threshold", best.method = "youden")
cat("最优截断值：", optimal_thresh, "\n")

4. 生成对应截断值的混淆矩阵 #

用最优阈值将预测概率转换为分类标签，再生成混淆矩阵：

# 基于最优阈值生成测试集预测标签
test_pred_label <- ifelse(test_pred_prob >= optimal_thresh, 1, 0)

# 基础R生成混淆矩阵
confusion_matrix <- table(真实标签 = test$asthma, 预测标签 = test_pred_label)
print(confusion_matrix)

# 若需更详细指标（可选，需加载caret包）
# library(caret)
# confusionMatrix(factor(test_pred_label), factor(test$asthma), positive = "1")

完整整合代码 #

将上述步骤补充到原有代码中，完整代码如下：

df<-data.frame(date = c(20120324,20120329,20121216,20130216,20130725,20130729,20130930,20131015,20131124,20131225,
                        20140324,20140530,20140613,20140721,20140630,20150102,20150214,20150312,20150316,20150329),
               temperature=c(35,36.5,34.3,37.8,39,40,34.5,35.9,35.8,36.1,37,35,36,36.3,37.8,38.1,39.2,34.5,34.9,35.2),
               bmi=c(20,23,25,27,32,24,35,21,19,29,21,32,21,22,24,25,19,18,25,26),
               asthma=c(1,1,0,1,0,0,1,1,0,0,1,1,1,0,1,0,0,1,1,0))

set.seed(101)
# 选取70%数据作为训练集
sample <- sample.int(n = nrow(df), size = floor(.7*nrow(df)), replace = F)
train <- df[sample, ]
test  <- df[-sample, ]

x<-model.matrix(asthma ~ temperature+bmi,data=train)[,-1]
y<-as.matrix(train$asthma)

library(glmnet)
glmmod <- glmnet(x, y, alpha=1, family="binomial")
cvfit = cv.glmnet(x, y)
coef_cv=coef(cvfit, s = "lambda.min")

# --- 新增步骤 ---
# 1. 生成预测概率
train_pred_prob <- predict(cvfit, newx = x, s = "lambda.min", type = "response")
x_test <- model.matrix(asthma ~ temperature+bmi, data = test)[,-1]
test_pred_prob <- predict(cvfit, newx = x_test, s = "lambda.min", type = "response")

# 2. 绘制预测概率分布图
library(ggplot2)
train_plot_data <- data.frame(
  true_label = factor(train$asthma, levels = c(0,1), labels = c("非哮喘", "哮喘")),
  pred_prob = as.vector(train_pred_prob)
)
ggplot(train_plot_data, aes(x = pred_prob, fill = true_label)) +
  geom_density(alpha = 0.5) +
  labs(x = "预测概率", y = "密度", title = "训练集预测概率分布", fill = "真实类别") +
  theme_minimal()

# 3. 寻找最优截断值
library(pROC)
roc_obj <- roc(test$asthma, test_pred_prob)
optimal_thresh <- coords(roc_obj, "best", ret = "threshold", best.method = "youden")
cat("最优截断值：", optimal_thresh, "\n")

# 4. 生成混淆矩阵
test_pred_label <- ifelse(test_pred_prob >= optimal_thresh, 1, 0)
confusion_matrix <- table(真实标签 = test$asthma, 预测标签 = test_pred_label)
print(confusion_matrix)

内容的提问来源于stack exchange，提问作者Lee