0

Does this sound legit, for people working with CNN and Time Series?

I have a Framework that applies Dynamic Tim Warping (DTW) on time series, using the DTW distance matrix, I cluster my data and create labels for each time series and then apply a 1DConvNet to make classification.

I end up with a very high validation accuracy, ca. 95%, and it also does not really look like overfitting or underfitting to me. But a 95% accuracy is too high in my eyes. I have a dataset that has 150 time series and each has 200 time steps. For clustering, I use KMeans Cluster with 4 clusters.

Accuracy

Loss

Skobo Do
  • 51
  • 6
  • 1
    Just a quick comment: I'd check the size of your validation split and eventually do a k-fold cross validation to be sure that you didn't randomly sample a 'too simple' validation set. What looks a bit weird is that the curve reaches 0.95 and after 1k steps then kind of stops there - that might be due to the selection of validation set. From what's visible, everything else looks legit to me though. – Chillston Oct 28 '22 at 17:59
  • @Chillston you mean this clean cut of the blue curve? Yeah that what I saw too. I'll try the k-fold validation. Thank you. – Skobo Do Oct 28 '22 at 21:03

0 Answers0