Fix training size when 'split_type'='time'

[ffernest]

Hi everyone,

I noticed that FLAML will incorporate the training data with the latest test set for the next fold during the CV of time-splitted data. Is there any setting I can fix the training size, i.e. adjust the training set start time when incorporate the new test set? It is the same behaviour of the max_train_size attribute in TimeSeriesSplit of sklearn. Thanks for any suggestions.

[sonichi]

You can pass a customized splitter to split_type

[ffernest]

I am just wondering if it can be customized. I will try it out. Thanks!

[ffernest]

I can directly assign the TimeSeriesSplit object to "split_type" and it behaves exactly as I want. Thanks for your help!