I have an $n$-dimensional time-series to apply LSTM to, $n$ is the number of features for each time point. These features can be clustered according to their concept, for example $n_1, ..., n_4$ are features related to let's say thermodynamic variables such as temperature and pressure, $n_5, ..., n_{10}$ are features related to geographical variables and so on.

So what would be the best practice for designing an LSTM topology for this problem? possible answers that comes to my mind are:

  • use primary many-to-one LSTMs for each cluster of features, and feed the outputs to a Dense network
  • use 1 single global LSTM (still many-to-$l$, with $l$ being the number of labels) for all of the features, and let the model decide if it wants to decouple the implicit clusters of features

So the first method is simpler, with less trainable parameters, but with hard-coded features, and the second one, not simple and more parameters but still more able to capture complexities


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.