如何将常规多变量DataFrame转换为Keras中LSTM适配的3D格式？

问题背景 #

在RStudio中用Keras构建LSTM模型，目标是用股票历史数据、新冠疫情数据等多特征预测股价。已完成数据归一化，训练/测试数据结构如下：

训练数据片段 #

> head(X_train)
       date  BNTX.Open  BNTX.High   BNTX.Low BNTX.Volume  VIX.Kurs
1 -1.726310 -0.9389774 -0.9147491 -0.9406800  -1.0391883 -1.044287
2 -1.718655 -0.9171379 -0.8772236 -0.9010828  -0.9745892 -1.060324
3 -1.710999 -0.8416828 -0.8386974 -0.8894619  -0.8806268 -1.065984
4 -1.703344 -0.8737693 -0.8639145 -0.8571816  -1.0318373 -1.098059
5 -1.695688 -0.8504807 -0.8601119 -0.8729989  -1.0149996 -1.183905
6 -1.688033 -0.8666275 -0.8773236 -0.9214194  -0.9598871 -1.182018
  New_deaths New_cases
1  -1.900206 -1.529246
2  -1.900206 -1.529233
3  -1.900206 -1.529246
4  -1.900206 -1.529246
5  -1.900206 -1.529246
6  -1.900206 -1.529246
> head(y_train)
  BNTX.Adjusted
1    -0.9144203
2    -0.8675775
3    -0.8803998
4    -0.8622004
5    -0.8811237
6    -0.9182463

当前模型代码 #

dim(X_train)#[1] 452   8

modelLSTM <- keras_model_sequential() %>% 
  layer_lstm(input_shape =  c(8, 1), units = 8, activation = "relu", return_sequences = FALSE) %>% 
  layer_dense(units = 8, activation = "relu") %>% 
  layer_dense(units = 1)
summary(modelLSTM)

modelLSTM %>% compile(loss = "mse", optimizer ="rmsprop", metrics = "mae")
modelLSTM <- modelLSTM %>% fit(X_train,y_train, epochs = 100, batch_size = 32, callback =list(callback_early_stopping(monitor = "val_loss", patience = 10, restore_best_weights = TRUE)) , validation_split = 0.2)

modelLSTM %>% evaluate(X_test, y_test)
pred_testy <- modelLSTM %>% predict(X_test)
mean((y_test-pred_testy)^2)
plot(y_test,pred_testy )

遇到的问题：直接设置input_shape后模型可能未正常工作；用生成器函数时无法区分特征集与目标集，无法完成测试集预测。

解决方案：多变量数据转3D张量 #

LSTM要求输入为3D张量：(样本数, 时间步长, 特征数)。核心是确定时间步长（用过去N个时间点的特征预测下一个时间点的股价），再将二维DataFrame转换为符合要求的3D结构。

1. 定义数据转换函数 #

该函数可明确区分特征集与目标集，完成训练/测试数据的格式转换：

create_lstm_dataset <- function(X, y, time_steps = 1) {
  # 初始化空列表存储转换后的数据
  X_data <- list()
  y_data <- list()
  
  # 遍历生成每个样本的时间序列窗口
  for (i in 1:(nrow(X) - time_steps)) {
    # 取连续time_steps行的特征作为一个样本
    X_window <- X[i:(i + time_steps - 1), ]
    # 取窗口下一个时间点的目标值
    y_target <- y[i + time_steps, ]
    
    X_data[[length(X_data) + 1]] <- as.matrix(X_window)
    y_data[[length(y_data) + 1]] <- as.matrix(y_target)
  }
  
  # 转换为3D张量：(样本数, 时间步长, 特征数)
  X_array <- array(unlist(X_data), dim = c(length(X_data), time_steps, ncol(X)))
  y_array <- array(unlist(y_data), dim = c(length(y_data), ncol(y)))
  
  return(list(X = X_array, y = y_array))
}

2. 转换训练和测试数据 #

假设选择时间步长为5（用过去5天的数据预测第6天的股价），执行转换：

# 设置时间步长
time_steps <- 5

# 转换训练数据
train_dataset <- create_lstm_dataset(X_train, y_train, time_steps)
X_train_lstm <- train_dataset$X
y_train_lstm <- train_dataset$y

# 转换测试数据
test_dataset <- create_lstm_dataset(X_test, y_test, time_steps)
X_test_lstm <- test_dataset$X
y_test_lstm <- test_dataset$y

# 查看转换后的数据维度
dim(X_train_lstm) # 输出应为 (452-5, 5, 8) = (447,5,8)
dim(y_train_lstm) # 输出应为 (447,1)

3. 调整LSTM模型输入 #

根据转换后的3D张量调整input_shape，格式为(时间步长, 特征数)：

modelLSTM <- keras_model_sequential() %>% 
  # input_shape = (time_steps, 特征数)，这里特征数是8
  layer_lstm(input_shape = c(time_steps, ncol(X_train)), units = 8, activation = "relu", return_sequences = FALSE) %>% 
  layer_dense(units = 8, activation = "relu") %>% 
  layer_dense(units = 1)

summary(modelLSTM)

4. 重新训练和评估模型 #

用转换后的3D数据训练模型，测试集同样使用转换后的结构：

modelLSTM %>% compile(loss = "mse", optimizer ="rmsprop", metrics = "mae")

# 训练模型
modelLSTM <- modelLSTM %>% fit(
  x = X_train_lstm, 
  y = y_train_lstm, 
  epochs = 100, 
  batch_size = 32, 
  callbacks = list(callback_early_stopping(monitor = "val_loss", patience = 10, restore_best_weights = TRUE)), 
  validation_split = 0.2
)

# 评估测试集
modelLSTM %>% evaluate(X_test_lstm, y_test_lstm)

# 生成预测
pred_testy <- modelLSTM %>% predict(X_test_lstm)
mean((y_test_lstm - pred_testy)^2)
plot(y_test_lstm, pred_testy)

关键说明 #

• 时间步长time_steps是超参数，可根据数据特点调整（如尝试3、5、7天等）
• 转换函数确保了特征窗口与目标值的对应关系：每个样本窗口对应下一个时间点的股价
• 转换后的3D张量完全符合Keras LSTM的输入要求，避免了维度不匹配问题

内容的提问来源于stack exchange，提问作者magNGHI