philipperemy/keras-tcn

Question: Multivariate TCN

farzinaa opened this issue · 4 comments

farzinaa commented

Hi,
I Have a question about the dependencies of the elements in an input sequence: Is there anyways to learn the relationship of such element?
For example in the exchange task there are 8 currencies. My understanding is that removing half of these currencies would not effect the prediction of the other four. OR adding extra features is redundant and will not effect the outcome of the prediction. Is this correct?
If so, is it possible for the model to learn from other currencies/features and how one would approach this?
Many thanks

loySoGxj commented

What kind of problem (background) are you thinking about exactly?

farzinaa commented

I am trying to predict time series with multiple features. The data are pressure in each valve in a system where valves are connected through pipes and pressure in valve A is always grater than pressure in valve B and pressure in valve C is lower than in valves B.
I use sequences of [A, B, C, ...]. However sometimes the predicted value of B is grater than value of A.

philipperemy commented

@farzinaa Any recurrent model in Keras expects a matrix of shape (batch_size, time_steps, input_dim).

Similar to LSTM, input_dim > 1 is the multivariate case.

The dependencies between each element in this dimension will be modeled.

For example, let's consider you have 10 time series of 50 steps. You have two choices:

  • Treat them independently (univariate case): your input will look like (batch_size=10, time_steps=50, input_dim=1).
  • Treat them jointly (multivariate case): your input will look like (batch_size=1, time_steps=50, input_dim=10).

I hope this answers your question.

farzinaa commented

Thank you. It is all clear now.