关联时序数据集降维：最优深度学习架构选型咨询

Hey there! Let's tackle your problem step by step.

Since your data is time-series with both inter-variable and temporal dependencies, the best fit here is an LSTM (Long Short-Term Memory) network. LSTMs are designed explicitly to capture long-term and short-term temporal patterns in sequential data, while naturally handling multi-variable inputs.

If you wanted a more compact alternative, you could also consider GRUs (Gated Recurrent Units), but LSTMs are more robust for capturing complex temporal relationships. For pure dimensionality reduction with sequence preservation, a time-series autoencoder (using LSTMs in encoder/decoder) works too—but since you're targeting a direct reduction to 1 dimension, an LSTM followed by a dense layer is simpler and more direct.

First, let's fix the input shape (Keras expects LSTM inputs in (batch_size, timesteps, features) format) and complete the model definition:

import numpy as np
from keras.models import Sequential
from keras.layers import LSTM, Dense

# Your original dataset: 100 timesteps, 10 variables
dataset = np.arange(1000).reshape(100, 10)

# Reshape for LSTM: add batch dimension (here, we treat the entire sequence as 1 batch)
X = dataset.reshape(1, 100, 10)

# Define the model
model = Sequential()
# Add LSTM layer: 32 hidden units (adjust based on your needs), input shape matches (timesteps, features)
model.add(LSTM(units=32, input_shape=(100, 10), return_sequences=False))
# Dense layer to reduce to 1 dimension
model.add(Dense(units=1))

# Optional: Compile the model if you plan to train it (you'll need a target y for training)
model.compile(optimizer='adam', loss='mse')

# Print model summary to verify
model.summary()

Key Notes: #

• We reshape X to (1, 100, 10) because Keras requires a batch dimension. If you had multiple independent sequences, you'd adjust the batch size accordingly.
• The units=32 in LSTM is a hyperparameter—you can tune this (try 16, 64, etc.) based on how much temporal information you need to capture.
• return_sequences=False means we only output the final hidden state of the LSTM (perfect for reducing the entire sequence to a single vector before the dense layer).
• If you have a target value to train against (e.g., a label for each sequence), you'd need to define a y array and use model.fit(X, y, ...) to train the model.

内容的提问来源于stack exchange，提问作者Roman