I have different time series sequences with varying lengths and want to use them as input for an LSTM. My approach is to use padding to fill shorter sequences.

Is there any best practice about which value to use for padding of time series? If I use zero padding and standardization as preprocessing, then the network would consider the padded entry as mean value. Would using the minimal value of the time series for the padded entries be more suitable?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.