I've heard about time-series classification being done with TCN's and CNN's combined with LSTM's very often, citing that CNN's would provide insight both forward and in the past since you already have all the information for that time period. For my application, there is a distinct shape and I'd like to classify whether it exists or not. For example, I want to detect whether the data looks like this enter image description here or this enter image description here

Of course, there would be noise involved and the feature would be much less obvious making the problem worthy of using machine learning. Is there some way I can exploit this knowledge of there being a single important feature (this hump) to use a different architecture or do anything differently?

  • $\begingroup$ Does this problem have only 1 input variable ? $\endgroup$ Mar 19 '19 at 6:58
  • $\begingroup$ Yes, as a function of time, but I would think the answer would apply to more complex problems? $\endgroup$ Mar 19 '19 at 20:04

This specific problem looks at the pattern across the whole data I.e. pattern will not show up from time < -3 or time > 3 for a given curvature.

You can try two models :

  1. Simple feed-forward Network with number of inputs = number of time steps (Maybe scale / shift the data so that it always has the same number of time steps )

This should be able to detect some patterns for classification (Like f(0) must be less that f(4))

  1. Univariate LSTM with different sizes of time steps in each sample

This should be able to learn that f(x) should stay near constant, reduce and then increase and return to constant

Both networks will have a sigmoid in output layer since it is a binary classification problem.

Code exmaple for LSTM : https://machinelearningmastery.com/sequence-classification-lstm-recurrent-neural-networks-python-keras/

  • $\begingroup$ How exactly would this take into account the existence of that central feature? Sorry I don't see exactly where this would be different from any standard time classification approach $\endgroup$ Mar 21 '19 at 5:08
  • $\begingroup$ Time classification works well with patterns that repeat (Say f(x) dips below 0 every N steps). This patterns happens only once. So , network have to learn parts of the pattern. $\endgroup$ Mar 21 '19 at 13:12

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.