I have a dataset of 90 periodic signals (current signals). These signals are divided into two big areas: Fault-free area and Fault area.

I sliced the signals using a window of 80ms with an overlap of 40ms in order to increase the dimensionality of the original dataset. Each window is labeled according to the part of the series which it was extracted from. Such approach generated a lot of points which are essentially the same due to the periodicity of the signal. As a consequence, when I try to fit my model and perform cross validation, I receive insane results such as 0.9999% accuracy. This is due to the fact that I have way too similar data points in my training set and in my validation set.

How can I solve this problem? Do you have any ideas?

  • $\begingroup$ How do you know that .99% accuracy is unreasonable? Maybe the labels are easy to predict. $\endgroup$ – Brian Spiering May 16 at 15:58
  • $\begingroup$ Because I used some visualization techniques (t-SNE) and I see that the data are not easily separable. Moreover, if I manually remove more than a half of the features I still have a insanely high validation score with a standard deviation of .00003 $\endgroup$ – Peppe Russo May 17 at 9:18

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.