I am working on a time series dataset in which each time step can be classified under 4 classes:

  • ~EOI : P(~EOI) = .85
  • EOI-1: P(EOI-1) = .05
  • EOI-2: P(EOI-2) = .05
  • EOI-3: P(EOI-3) = .05

where EOI is Event Of Interest.

Due to the huge class imbalance in ~EOI and EOI-X, I planned to use the boosting technique and build 1 classifier to identify EOI/~EOI events and then a sub-class classifier which would classify the 3 EOI events.

On further inspection I fear that for ~EOI class, there may be too much randomness and the may not have any distinguishable pattern and may just be noisy data.

I tried the Classification approach(traditional ML + NN approach), but i feel that could be set up for failure.

Another approach

1 approach that I could think of was to apply an auto encoder to do Anomaly detection. But before I invest time and effort there, i needed some opinion/advice on the same.

Please suggest alternative approaches or possible pros/cons of my current approach

  • $\begingroup$ This looks complicated, have you tried something simpler first? $\endgroup$ Sep 26, 2023 at 7:37
  • $\begingroup$ what could be simpler?? I mean i did try binary classification $\endgroup$ Sep 26, 2023 at 7:58


Your Answer

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