5 SVM在R中的实现

Author

Li Zongzhang

Published

October 17, 2025

本章介绍R中的SVM

安装包

install.packages("tidyverse")
install.packages("MASS")
install.packages("klaR")
install.packages("devtools")
install.packages("psych")
install.packages("MVN")
install.packages("e1071")

加载包

library(tidyverse)
library(psych)
library(biotools)
library(MVN)
library(e1071)

1 SVM模型的建立

# 取两个变量做演示
iris_sub <- iris[, c("Sepal.Length", "Sepal.Width", "Species")]

# 建立 SVM 模型（径向基核函数 RBF）
svm_model <- svm(Species ~ ., data = iris_sub, 
                 kernel = "radial", 
                 cost = 1, 
                 gamma = 0.5)

2 SVM模型的可视化

# 生成网格点用于预测
xrange <- seq(min(iris_sub$Sepal.Length) - 0.5, 
              max(iris_sub$Sepal.Length) + 0.5, 
              by = 0.02)

yrange <- seq(min(iris_sub$Sepal.Width) - 0.5, 
              max(iris_sub$Sepal.Width) + 0.5, 
              by = 0.02)

grid <- expand.grid(Sepal.Length = xrange, 
                    Sepal.Width = yrange)

# 对网格点分类预测
grid$Species <- predict(svm_model, grid)


# 画出分类区域和样本点
ggplot() +
  geom_tile(data = grid, 
            aes(x = Sepal.Length, 
                y = Sepal.Width, 
                fill = Species), 
            alpha = 0.3) +
  geom_point(data = iris_sub, 
             aes(x = Sepal.Length, 
                 y = Sepal.Width, 
                 color = Species), 
             size = 2) +
  labs(title = "SVM on Iris Dataset (3-class)",
       x = "Sepal Length", 
       y = "Sepal Width") +
  theme_minimal()

3 计算预测准确率

pred <- predict(svm_model, iris_sub)

accuracy <- mean(pred == iris_sub$Species)

print(paste("Accuracy:", round(accuracy * 100, 2), "%"))

[1] "Accuracy: 82 %"

4 改进SVM模型

4.1 调整参数

svm_model_tuned <- svm(Species ~ ., 
                       data = iris_sub,
                       kernel = "radial", 
                       cost = 10, 
                       gamma = 0.8)

pred_tuned <- predict(svm_model_tuned, iris_sub)

accuracy_tuned <- mean(pred_tuned == iris_sub$Species)

print(paste("Tuned Accuracy:", round(accuracy_tuned * 100, 2), "%"))

[1] "Tuned Accuracy: 82 %"

4.2 使用不同核函数（线性核）

svm_model_linear <- svm(Species ~ ., 
                        data = iris_sub,
                        kernel = "linear", 
                        cost = 10)

pred_linear <- predict(svm_model_linear, 
                       iris_sub)

accuracy_linear <- mean(pred_linear == iris_sub$Species)

print(paste("Linear Kernel Accuracy:", round(accuracy_linear * 100, 2),
            "%"))

[1] "Linear Kernel Accuracy: 82 %"

4.3 数据预处理（标准化）

iris_sub_scaled <- iris_sub

iris_sub_scaled[, 1:2] <- scale(iris_sub_scaled[, 1:2])

svm_model_scaled <- svm(Species ~ ., 
                        data = iris_sub_scaled,
                        kernel = "radial", 
                        cost = 1, 
                        gamma = 0.5)

pred_scaled <- predict(svm_model_scaled, iris_sub_scaled)

accuracy_scaled <- mean(pred_scaled == iris_sub_scaled$Species)

print(paste("Scaled Data Accuracy:", round(accuracy_scaled * 100, 2),
            "%"))

[1] "Scaled Data Accuracy: 82 %"

4.4 交叉验证选择最佳参数

tune_result <- tune(svm, Species ~ ., 
                    data = iris_sub,
                    ranges = list(cost = 10^(-1:2), 
                                  gamma = c(0.1, 0.5, 1)))

best_model <- tune_result$best.model

pred_best <- predict(best_model, iris_sub)

accuracy_best <- mean(pred_best == iris_sub$Species)

print(paste("Best Model Accuracy:", round(accuracy_best * 100, 2),
            "%"))

[1] "Best Model Accuracy: 80 %"

# 输出最佳参数
print(tune_result$best.parameters)

  cost gamma
5  0.1   0.5

# 输出调参结果
print(tune_result)


Parameter tuning of 'svm':

- sampling method: 10-fold cross validation 

- best parameters:
 cost gamma
  0.1   0.5

- best performance: 0.2133333

# 画出最佳模型的分类区域和样本点
grid$Species <- predict(best_model, grid)
ggplot() +
  geom_tile(data = grid, 
            aes(x = Sepal.Length, 
                y = Sepal.Width, 
                fill = Species), 
            alpha = 0.3) +
  geom_point(data = iris_sub, 
             aes(x = Sepal.Length, 
                 y = Sepal.Width, 
                 color = Species), 
             size = 2) +
  labs(title = "Tuned SVM on Iris Dataset (3-class)",
       x = "Sepal Length", y = "Sepal Width") +
  theme_minimal()

4.5. 使用更多特征

# 前文仅用两个特征做可视化，实际可用全部特征
# 由于可视化限制，无法直接展示多维特征空间的分类边界，但可以通过准确率来评估模型性能

svm_model_full <- svm(Species ~ ., 
                      data = iris,
                      kernel = "radial", 
                      cost = 1, 
                      gamma = 0.5)

pred_full <- predict(svm_model_full, iris)

accuracy_full <- mean(pred_full == iris$Species)

print(paste("Full Feature Set Accuracy:", 
            round(accuracy_full * 100, 2), "%"))

[1] "Full Feature Set Accuracy: 97.33 %"