使用Categorical Crossentropy后LSTM测试准确率仍为0的问题求助

我尝试基于visit_date（日期类型）预测visit_count（整数类型），但测试准确率始终为0.0，训练结果如下：

输入数据样例 #

完整代码 #

# seperating the date and count
date = data3['visit_date']
count = data3['visit_count']

labels = pd.get_dummies(date)

X_train, X_test, y_train, y_test = train_test_split(count, labels, test_size=0.2, random_state=42)
# Split the data into training and testing sets

# Define the LSTM model architecture
embedding_dim = 100
model = Sequential()
model.add(Embedding(100, embedding_dim, input_length=100))
model.add(SpatialDropout1D(0.2))
model.add(LSTM(100, dropout=0.2, recurrent_dropout=0.2))
model.add(Dense(len(labels.columns), activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

# Train the LSTM model
epochs = 100
batch_size = 640
history = model.fit(X_train, y_train, epochs=epochs, batch_size=batch_size, validation_data=(X_test, y_test), verbose=2)

loss, accuracy = model.evaluate(X_test, y_test, verbose=0)
print("LSTM Test Accuracy:", accuracy)

尽管使用了categorical_crossentropy作为损失函数，LSTM模型的测试准确率仍为0，需要解决该问题。

问题根源分析 #

1. 任务类型完全错误：你要做的是回归任务（预测整数型访问量），但代码按多分类任务构建：
   • 用pd.get_dummies(date)把日期转成独热编码当标签，实际标签应该是visit_count
   • categorical_crossentropy损失和accuracy指标都是分类任务的配置，完全不匹配回归场景
2. 输入输出颠倒：代码把count（访问量）作为模型输入，date的独热编码作为标签，和「用日期预测访问量」的目标完全相反
3. LSTM输入格式错误：LSTM要求输入是三维张量(样本数, 时间步长, 特征数)，但当前输入是一维的count，且Embedding层参数完全不符合日期特征的处理逻辑

修正方案 #

1. 调整任务定位与数据处理 #

把任务回归化，正确划分输入（日期特征）和标签（访问量），先将日期转成数值型特征：

import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense

# 日期转datetime格式并提取特征
data3['visit_date'] = pd.to_datetime(data3['visit_date'])
data3['year'] = data3['visit_date'].dt.year
data3['month'] = data3['visit_date'].dt.month
data3['day'] = data3['visit_date'].dt.day
data3['weekday'] = data3['visit_date'].dt.weekday

# 定义输入特征和标签
X = data3[['year', 'month', 'day', 'weekday']].values
y = data3['visit_count'].values

# 划分训练测试集
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# 调整LSTM输入格式为三维：(样本数, 时间步长, 特征数)，这里时间步长设为1
X_train = X_train.reshape((X_train.shape[0], 1, X_train.shape[1]))
X_test = X_test.reshape((X_test.shape[0], 1, X_test.shape[1]))

2. 构建回归型LSTM模型 #

# 定义回归模型，输出单个数值
model = Sequential()
model.add(LSTM(50, input_shape=(X_train.shape[1], X_train.shape[2])))
model.add(Dense(1))
# 回归任务用均方误差(MSE)作为损失，平均绝对误差(MAE)作为评估指标
model.compile(optimizer='adam', loss='mse', metrics=['mae'])

# 训练模型
history = model.fit(X_train, y_train, epochs=50, batch_size=32, validation_data=(X_test, y_test), verbose=2)

# 评估模型
loss, mae = model.evaluate(X_test, y_test, verbose=0)
print("LSTM Test MAE:", mae)

3. 进阶优化：时序滑动窗口 #

如果是用过去N天的数据预测下一天访问量，可构建滑动窗口数据集：

def create_dataset(X, y, time_steps=3):
    Xs, ys = [], []
    for i in range(len(X) - time_steps):
        # 取过去time_steps天的特征
        Xs.append(X[i:(i + time_steps)])
        # 取第time_steps+1天的访问量作为标签
        ys.append(y[i + time_steps])
    return np.array(Xs), np.array(ys)

# 用过去3天的特征预测第4天的访问量
X_seq, y_seq = create_dataset(X, y, time_steps=3)
X_train_seq, X_test_seq, y_train_seq, y_test_seq = train_test_split(X_seq, y_seq, test_size=0.2, random_state=42)

# 构建模型
model = Sequential()
model.add(LSTM(50, input_shape=(X_train_seq.shape[1], X_train_seq.shape[2])))
model.add(Dense(1))
model.compile(optimizer='adam', loss='mse')

history = model.fit(X_train_seq, y_train_seq, epochs=50, batch_size=32, validation_data=(X_test_seq, y_test_seq), verbose=2)

内容的提问来源于stack exchange，提问作者Sonny