TimeSeriesSplit for big unbalanced data set: Training data only contain information about one class

Question

I ran into an issue when doing cross-validation with TimeSeriesSplit on my big unbalanced time-series data set

As my Y is 98.1% composed of the same class, when doing the TimeSeriesSplit it is not possible to guarantee that in every interaction I have 2 or more classes.

ValueError: Training data only contain information about one class.

That is, the frequency of values different from this majority class in Y is so low that when doing a cross validation on a rolling basis, as shown in the image below, I don't have enough classes to cross-validate.

Are there any other suggestions for me to do the validation process of a strongly unbalanced time series? Additionally, as it is a time series, I cannot use k-fold based cross-validation, for example.

Does this answer your question? Using k-fold cross-validation for time-series model selection — Adrian Keister, Oct 07 '22 at 18:37
@AdrianKeister, actually this does not answer my question, since it doesn't address a methodology to do cross-validation on an unbalanced time series, which is exactly what I'm looking for. That is, I need a way to cross-validate and ensure that there is always more than one class in each of the interactions that the cross-validation process will generate. — Tazz, Oct 07 '22 at 18:47
Is your data time-series? If yes, it seems like you just need to predict the brown class, as it is the only one in the most recent data. — Tim, Oct 07 '22 at 19:26

TimeSeriesSplit for big unbalanced data set: Training data only contain information about one class

0 Answers0