蓝莓产量预测（R语言版）

数据描述

字段名	描述	字段名	描述
id	蓝莓唯一标识	MinOfUpperTRange	花期内最高温带日平均气温的最低记录,
Clonesize	蓝莓克隆平均大小	AverageOfUpperTRange	花期内最高温带日平均气温,
Honeybee	蜜蜂密度	MaxOfLowerTRange	花期内最低温带日平均气温的最高记录,
Bumbles	大型蜜蜂密度	MinOfLowerTRange	花期内最低温带日平均气温的最低记录,
Andrena	安德烈纳蜂密度	AverageOfLowerTRange	花期内最低温带日平均气温,
Osmia	钥匙蜂密度	RainingDays	花期内降雨量大于0的日数总和,
MaxOfUpperTRange	花期内最高温带日平均气温的最高记录,	AverageRainingDays	花期内降雨日数的平均值,
fruitset	果实集	seeds	种子数
fruitmass	果实质量	yield	产量

数据预处理

# 读取数据
train_data <- read.csv("D:\\大三上\\r语言\\期末\\train.csv")
cat('数据集信息：\n')
str(train_data)
summary_data <- as.data.frame(summary(train_data))
summary_data<-t(summary_data)
# 显示数据框
print(summary_data)
summary(train_data)
# 查看各列缺失值
cat('数据集信息缺失情况：\n')
print(colSums(is.na(train_data)))

#将train_data数据集中有缺失数据所在行删掉
train_data<-train_data[complete.cases(train_data$honeybee, train_data$bumbles,train_data$MaxOfUpperTRange,
                                        train_data$MaxOfLowerTRange,train_data$MinOfLowerTRange), , drop = FALSE]
#再次检验缺失值
print(colSums(is.na(train_data)))

# 查看重复值
cat('数据集信息重复情况：\n')
print(sum(duplicated(train_data)))
cat(rep('-', 15), '\n')
set.seed(123)  # Set seed for reproducibility

#install.packages("corrplot")
library(corrplot)
col<-cor(train_data)
# 设置整体图形的大小
par(mar = c(1.2, 1.2, 1.2, 1.2))
corrplot(col, method = "color", addCoef.col = "black", tl.cex = 0.8,number.cex = 0.5)

par(mar = c(3.0,3.0,2.0,2.0))
hist(train_data$AverageOfUpperTRange,freq = FALSE)
lines(density(train_data$AverageOfUpperTRange),col='blue')
rug(jitter(train_data$AverageOfUpperTRange))


# 导入必要的库
library(ggplot2)
# 绘制yield属性的盒图
ggplot(data = train_data, aes(x = yield)) +
  geom_boxplot(fill = "lightblue") +
  geom_boxplot(fill = "blue", outlier.shape = NA, coef = 1.5, width = 0.2) +  # 设置填充颜色为蓝色，移除离群值的标记，调整箱体宽度
  theme_minimal() +
  ggtitle("Boxplot of Yield")+
  theme(plot.title = element_text(hjust = 0.5))

# 导入必要的库
library(ggplot2)

# 绘制yield属性的盒图
ggplot(data = train_data, aes(x = yield)) +
  geom_boxplot(fill = "lightblue") +
  geom_boxplot(fill = "blue", outlier.shape = NA, coef = 1.5, width = 0.2) +  # 设置填充颜色为蓝色，移除离群值的标记，调整箱体宽度
  theme_minimal() +
  ggtitle("Boxplot of Yield")+
  theme(plot.title = element_text(hjust = 0.5))

#按数据集的分类特征分布
# 属性分布箱线图
library(reshape2)
# 选择分类特征列
nominal_df <- train_data[, c('MaxOfUpperTRange', 'MinOfUpperTRange', 'AverageOfUpperTRange',
                     'MaxOfLowerTRange', 'MinOfLowerTRange', 'AverageOfLowerTRange',
                     'RainingDays', 'AverageRainingDays','yield')]


melted_df <- melt(nominal_df, id.vars = NULL)

# Create boxplot
p <- ggplot(melted_df, aes(x = variable, y = value, fill = as.factor(variable))) +
  geom_boxplot(fill="lightblue") +
  facet_wrap(~variable, scales = "free") +
  theme_minimal() +
  labs(x = "", y = "yield")

# Print the plot
print(p)

#数据集中蜜蜂类型的分布

hist_bumbles <- ggplot(train_data, aes(x = bumbles)) + 
  geom_histogram(fill="green") +
  ggtitle("Histogram of bumbles column")

hist_andrena <- ggplot(train_data, aes(x = andrena)) + 
  geom_histogram(fill="red") +
  ggtitle("Histogram of andrena column")

hist_osmia <- ggplot(train_data, aes(x = osmia)) + 
  geom_histogram(fill="yellow") +
  ggtitle("Histogram of osmia column")

hist_clonesize <- ggplot(train_data, aes(x = clonesize)) + 
  geom_histogram(fill="purple") +
  ggtitle("Histogram of clonesize column")

hist_honeybee <- ggplot(train_data, aes(x = honeybee)) + 
  geom_histogram(fill="pink") +
  ggtitle("Histogram of honeybee column")

# Arrange histograms in a grid
grid.arrange(hist_bumbles, hist_andrena, hist_osmia, hist_clonesize, hist_honeybee, ncol = 3)

# 加载 corrplot 库
library(corrplot)
# 画矩阵相关性图
corrplot(col, method = "color", addCoef.col = "black", tl.cex = 0.8, number.cex = 0.5)

# 设置图形边距
par(mar = c(3.0, 3.0, 2.0, 2.0))

# 绘制直方图
hist(train_data$honeybee, freq = FALSE, col = "lightblue", main = "Histogram and Density Plot", breaks = seq(min(train_data$honeybee), max(train_data$honeybee), by = 0.05)

# 绘制核密度估计曲线
lines(density(train_data$honeybee), col = 'blue')

# 调整 jitter 大小，增加数据点密度
rug(jitter(train_data$honeybee, amount = 0.01), col = "darkred", lwd = 1.5)

多元线性回归

检测多重线性

# 数据读取
train_data <- read.csv("D:\\学\\R作业\\大作业\\train.csv")  
x <- train_data[, !(names(train_data) %in% c("yield"))]

# 计算VIF
lm_model <- lm(x[,1] ~ ., data = x)  
vif_result <- car::vif(lm_model)

# 使用 kable 函数美化输出
kable(data.frame(Variable = names(vif_result), VIF = vif_result), format = "html", caption = "VIF Results") %>%
  kable_styling(full_width = FALSE)

主成分分析

# 进行主成分分析
pca_result <- prcomp(x, scale. = TRUE)  

# 计算主成分方差贡献率和累计方差贡献率
variance_contrib <- pca_result$sdev^2 / sum(pca_result$sdev^2)
cumulative_var_contrib <- cumsum(variance_contrib)

# 找到累积方差贡献率达到95%的主成分数量
num_components_95 <- which(cumulative_var_contrib >= 0.95)[1]

# 输出主成分方差贡献率和累计方差贡献率
print(data.frame(
  Principal_Component = 1:length(variance_contrib),
  Variance_Contribution = variance_contrib,
  Cumulative_Variance_Contribution = cumulative_var_contrib
))

# 输出累积方差贡献率达到95%的主成分数量和对应的主成分
cat("Number of components for 95% cumulative variance contribution:", num_components_95, "\n")
cat("Principal components for 95% cumulative variance contribution:", paste(1:num_components_95, collapse = ", "), "\n")

 建立多元线性回归模型

# 选择累计方差贡献率达到95%以上的主成分
selected_components <- pca_result$x[, 1:num_components_95]

# 合并主成分和目标变量，并转换为数据框
data_for_regression <- data.frame(cbind(selected_components, yield = train_data$yield))

# 建立多元线性回归模型
linear_model <- lm(yield ~ ., data = data_for_regression)

# 主成分分析摘要
print("Principal Component Analysis:")
kable(data.frame(
  Principal_Component = 1:length(variance_contrib),
  Variance_Contribution = variance_contrib,
  Cumulative_Variance_Contribution = cumulative_var_contrib
), format = "html", caption = "Principal Component Analysis") %>%
  kable_styling(full_width = FALSE)

# 输出主成分方差贡献率达到95%的主成分数量和对应的主成分
cat("\nNumber of components for 95% cumulative variance contribution:", num_components_95, "\n")
cat("Principal components for 95% cumulative variance contribution:", paste(1:num_components_95, collapse = ", "), "\n")

# 多元线性回归模型摘要
print("\nMultiple Linear Regression Model:")
model_summary <- summary(linear_model)

# 使用 kable 函数美化输出
kable(as.data.frame(model_summary$coefficients), format = "html", caption = "Multiple Linear Regression Model") %>%
  kable_styling(full_width = FALSE)

# 计算预测值
predicted_values <- fitted(linear_model)

# 计算残差
residuals <- residuals(linear_model)

# 计算MSE
mse <- mean(residuals^2)

# 计算R-squared
r_squared <- model_summary$r.squared

# 打印MSE和R-squared
cat("Mean Squared Error (MSE):", mse, "\n")
cat("R-squared (R2):", r_squared, "\n")

绘图检验

# 创建散点图（美化版）
scatter_plot <- ggplot(data = data_for_regression, aes(x = yield, y = predicted_values)) +
  geom_point(color = "blue", size = 0.5, alpha = 0.7) +  # 调整颜色、点的大小和透明度
  geom_smooth(method = "lm", se = FALSE, color = "red", linetype = "dashed") +
  labs(title = "Scatter Plot of Predicted vs Actual Yield",
       x = "Actual Yield",
       y = "Predicted Yield") +
  theme_minimal() +  # 使用简洁主题
  theme(legend.position = "none")  # 隐藏图例

# 打印美化散点图
print(scatter_plot)

###残差序列图
代码：
# 计算LOESS平滑曲线
smoothed_residuals <- loess.smooth(fitted(linear_model), residuals, span = 0.8)$y

# 绘制残差序列图（Residuals vs Fitted、Scale-Location、Residuals vs Leverage、Cook's Distance）
par(mfrow = c(2, 2))

# Residuals vs Fitted
plot(fitted(linear_model), residuals, main = "Residuals vs Fitted", xlab = "Fitted Values", ylab = "Residuals", col = "darkgreen", pch = 16, cex = 0.7)
lines(fitted(linear_model), smoothed_residuals, col = "red", lwd = 2)

# Scale-Location
sqrt_abs_residuals <- sqrt(abs(residuals))
plot(fitted(linear_model), sqrt_abs_residuals, main = "Scale-Location", xlab = "Fitted Values", ylab = "sqrt(|Residuals|)", col = "darkblue", pch = 16, cex = 0.7)
lines(fitted(linear_model), loess.smooth(fitted(linear_model), sqrt_abs_residuals, span = 0.8)$y, col = "red", lwd = 2)

# Residuals vs Leverage
plot(hatvalues(linear_model), residuals, main = "Residuals vs Leverage", xlab = "Leverage", ylab = "Residuals", col = "purple", pch = 16, cex = 0.7)
abline(h = 0, col = "red", lty = 2)

# Cook's Distance
cooksd <- cooks.distance(linear_model)
plot(cooksd, pch = "18", main = "Cook's Distance", col = "darkorange", xlab = "Obs Number", ylab = "Cook's distance", cex = 0.7)
abline(h = 4/(length(residuals) - length(coefficients(linear_model))), col = "red", lty = 2)

# 重置绘图参数
par(mfrow = c(1, 1))

###残差直方图

代码：
# 残差的直方图
hist(residuals, main = "Histogram of Residuals", col = "lightblue", border = "black", probability = TRUE)
lines(density(residuals), col = "red", lwd = 2)

###Q-Q图

代码：
# 绘制Q-Q图
qqnorm(residuals, main = "Q-Q Plot of Residuals", col = "blue")
qqline(residuals, col = "red")


###散点图

# 计算 LOESS 平滑曲线（调整 span 参数）
loess_fit <- loess(residuals ~ data_for_regression$yield, span = 0.7)  # 适当调整 span 的值

# 绘制散点图
plot(data_for_regression$yield, residuals, main = "Residuals vs. Actual Values", xlab = "Actual Values", ylab = "Residuals", col = "lightgreen", pch = 16)

# 添加 LOESS 拟合曲线
lines(data_for_regression$yield, predict(loess_fit), col = "red", lwd = 2)

##残差值与预测值

### 残差与预测值的散点图
# 计算LOESS平滑曲线
loess_fit <- loess(residuals ~ fitted(linear_model), span = 0.8)

# 绘制残差与拟合值的散点图
plot(fitted(linear_model), residuals, main = "Residuals vs. Fitted Values", xlab = "Fitted Values", ylab = "Residuals", col = "darkgreen", pch = 16)

# 添加LOESS拟合曲线
lines(fitted(linear_model), predict(loess_fit), col = "red", lwd = 2)

# 重置绘图参数
par(mfrow = c(1, 1))

随机森林

x <- train_data[, !(names(train_data) %in% c("yield"))]  # 选择除了"yield"列之外的所有列作为特征
y <- train_data$yield  # "yield"列作为目标变量
library(lattice)
library(caret)
# 使用caret包中的createDataPartition函数划分数据
index <- createDataPartition(y, p = 0.7, list = FALSE)
x_train <- x[index, ]
x_test <- x[-index, ]
y_train <- y[index]
y_test <- y[-index]
library(ranger)
# 模型建立
rf_model <- ranger(y_train ~ ., data = x_train, num.trees = 500)
rf_model <- ranger(y_train ~ ., data = x_train, num.trees = 500, importance = "impurity")

# 预测
y_pred <- predict(rf_model, data = x_test)$predictions
mse <- mean((y_test - y_pred)^2)
r2 <- 1 - mse / var(y_test)

#install.packages("ranger")
library(ranger)
library(caret)
library(lattice)
# 设置随机搜索
# 设置随机搜索
set.seed(17)
rf_grid <- expand.grid(
  mtry = c(1, 17, by=1),
  splitrule = c("variance"),
  min.node.size = c(2, 5, 10)
)

ctrl <- trainControl(method = "cv", number = 15)
rf_search <- train(x_train, y_train, method = "ranger", 
                   trControl = ctrl, tuneGrid = rf_grid)
rf_search_model <- rf_search$finalModel

# 获取最佳参数和评分
best_params <- rf_search$bestTune

#install.packages("ggplot2")
library(ggplot2)

# 提取交叉验证结果
cv_results <- rf_search$results
names(cv_results)
# 绘制超参数与性能之间的关系图
ggplot(cv_results, aes(x = mtry, y = RMSE)) +
  geom_point(size = 3) +
  labs(x = "mtry", y = "RMSE", 
       title = "Hyperparameter Tuning with Random Forest") +
  theme_minimal()


# 获取特征重要性
feature_importances <- ranger::importance(rf_model)

# 将命名向量转换为数据框
feature_importances_df <- data.frame(
  Feature = names(feature_importances),
  Importance = as.numeric(feature_importances)
)

# 按重要性降序排序
feature_importances_df <- feature_importances_df[order(-feature_importances_df$Importance), ]

# 打印特征重要性的DataFrame
print(feature_importances_df)


# 模型建立
rf_model <- ranger(y_train ~ ., data = x_train, num.trees = 500)
rf_model <- ranger(y_train ~ ., data = x_train, num.trees = 500, importance = "impurity")


# 使用训练好的模型对测试数据进行预测
test_data
test_predictions_rf <- predict(rf_model, data = x_test)$predictions


# 计算残差
residuals_rf <- test_predictions_rf - test_data$yield
class(residuals_rf)

mse <- mean((y_test - test_predictions_rf)^2)
r2 <- 1 - mse / var(y_test)

# 检查残差中的空值
missing_residuals <- which(is.na(residuals_rf))

# 移除残差中的空值
clean_residuals <- na.omit(residuals_rf)
# 计算残差的标准差
residual_sd <- sd(clean_residuals)
# 计算残差标准误差
n <- length(clean_residuals)
residual_se <- residual_sd / sqrt(n)

cat("Mean Squared Error (MSE):", mse, "\n")
cat("R-squared (R2):", r2, "\n")
#cat("Residuals:", residuals_rf, "\n")
cat("Residual Standard Error(RSE):", residual_se, "\n")

# 创建包含预测结果的新数据框
rf_test_data_with_predictions <- data.frame(x_test)
rf_test_data_with_predictions$predicted_yield <- test_predictions_rf

# 打印包含预测结果的数据框的前几行
print(head(rf_test_data_with_predictions))
#----------------------------由于randomForest运行时间太久，因此未采用以下代码——————————————————————————#

# 读取数据
train_data <- read.csv("D:\\学\\R作业\\大作业\\train.csv")

train_sub=sample(nrow(train_data),0.7*nrow(train_data))
train_data=train_data[train_sub,]
test_data=train_data[-train_sub,]
#按照7:3划分数据集

n<-length(names(train_data))    
#计算数据集中自变量个数
rate=1     
#设置模型误判率向量初始值
for(i in 1:(n-1)){
  set.seed(1234)
  rf_train<-randomForest(train_data$Label~.,data=train_data,mtry=i,ntree=1000)
  rate[i]<-mean(rf_train$err.rate)#计算基于OOB数据的模型误判率均值
}
rate     
#展示所有模型误判率的均值
plot(rate,type='b',main="不同mtry取值的误判率",
     xlab="n",ylab="err.rate")
mtry <- which.min(rate)
#mtry取误差率最低时的n

set.seed(100)
rf_train<-randomForest(train_data$yield~.,
                       data=train_data,mtry=mtry,
                       ntree=1000)
plot(rf_train, panel.first=grid(10, 10),main="模型误差与ntree关系")    
#绘制模型误差与决策树数量关系图  
#黑线代表决策树的error,另外两条是bagging后的error
ntree=600

rfm<-randomForest(yield~.,data=train_data,
                  importance=TRUE,proximity=TRUE,
                  mtry=mtry,ntree=ntree)
rfm
#install.packages("caret")
library(ggplot2)
library(lattice)
#install.packages("future.apply")
library(caret)
#install.packages("pROC")
library(pROC) #绘制ROC曲线

rf_test <- predict(rfm,newdata=test_data,type="class")
#在训练集上使用模型
rf_cf <- caret::confusionMatrix(as.factor(rf_test),test_data$yield)
#输出模型的相关评价指标
rf_cf

rf_roc <- roc(test_data$yield,as.numeric(rf_test))
plot(rf_roc, print.auc=TRUE, 
     auc.polygon=TRUE, grid=c(0.1, 0.2),
     grid.col="grey", max.auc.polygon=TRUE,
     auc.polygon.col="darkseagreen1", print.thres=TRUE,
     main='随机森林模型ROC曲线,mtry=4,ntree=200')
#绘制roc图

info=rfm$importance
info
varImpPlot(rfm, main = "衡量变量重要性的两个指标")
#MeanDecreaseAccuracy变量替换后随机森林预测准确性的降低程度
#MeanDecreaseGini变量替换后GINI系数（悬殊差距、异质性）的降低程度

data.frame(info)
importance=info[,3]+info[,4]
barplot(importance,cex.lab=0.5,main="各变量的重要性",col="darkseagreen")

 

模型对比

模型评估指标	MSE	R2	RSE
多元线性回归	381794.3	0.7864212	618.3
随机森林	310582.4	0.8273324	73.02739