I want to use CNN architectures for classification of multivariate time-series, where we apply one label to each sequence. I searched the net for the available designs in the literature and i found the following options. Assuming each MV time-series $\mathbf{S}$ is defined in $\mathcal{R}^{n \times t}$, where n is the number of dimensions and $t$ is the length of the sequence:

  1. Treating each $\mathbf{S}$ as a $n \times T$ image as the input of CNN.
  2. Use $n$ separate input channels of size $1 \times T$ and apply a CNN with 1D conv-filters for such input.
  3. Using $n$ separate CNNs each of which is associated with one dimension of $\mathbf{S}$ and combine their extracted features as the inputs of the fully connected layer.

So, I was wondering if there is any more efficient architecture suggested for such type of data

Also, as usually there an inconsistency between the sequence length of different time-series, is there any smart solution suggested for that?

  • $\begingroup$ The question was a partial answer for me! $\endgroup$
    – keiv.fly
    Aug 1, 2018 at 21:18


Your Answer

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