I have the following task to do: Training by the consecutive 3 days to predict the 4th day. Each day data represents one CSV file which has dimension 24x25. Every datapoints of each CSV file are pixels. I have required to use the models like regression(linear, ridge) and LSTM.
For training by 3 days for each model:
For regression model: After did the each day data by flattening, I transpose the data with Shape: (600, 3).
For lstm model: After did the each day data by flattening, I keep the data as it is with Shape: (3, 600).
For example like this:
day_1 = [0.1, 0.2, ..., 0.6] # 600 features for Day 1
day_2 = [0.15, 0.25, ..., 0.65] # 600 features for Day 2
day_3 = [0.2, 0.3, ..., 0.7] # 600 features for Day 3
X_train_linear_ridge = np.array([
[0.1, 0.15, 0.2], # Feature 1 across Day 1, Day 2, Day 3
[0.2, 0.25, 0.3], # Feature 2 across Day 1, Day 2, Day 3
# ...
[0.6, 0.65, 0.7] # Feature 600 across Day 1, Day 2, Day 3
]) # Shape: (600, 3)
X_train_lstm = np.array([
[0.1, 0.2, ..., 0.6], # Day 1 features
[0.15, 0.25, ..., 0.65], # Day 2 features
[0.2, 0.3, ..., 0.7] # Day 3 features
]) # Shape: (3, 600)
Would anyone tell me preparation of data for regression models with Shape: (600, 3) and lstm with Shape: (3, 600) are conceptually correct way?
My motivation:
LSTM: LSTMs are designed to process data with a sequential, temporal relationship. By feeding data with shape (3, 600) (representing 3 time steps, each with 600 features), the LSTM can learn patterns across the sequence. Each time step corresponds to a day in the data, while the 600 features represent individual values for that day. This structure is essential for the LSTM to leverage temporal dependencies.
Linear and Ridge Regression: These models lack the inherent sequential processing capability. They interpret each input as a single flat vector. To approximate sequential learning, we can treat each day's data as a separate feature, creating a setup with shape (600, 3), where the 600 features are "stacked" over the 3 days. Each day becomes a feature for the regression model, but it cannot capture temporal dependencies as an LSTM does.
My concept is correct for regressions and lstm model with respect to shape?
