autogluon/autogluon

Support for custom cross-validation strategies, e.g., forward chaining for temporally ordered tabular data

Aperta

#4492 aperta il 25 set 2024

 (4 commenti) (3 reazioni) (0 assegnatari)Python (1176 fork)batch import
enhancementhelp wantedmodule: tabular

Metriche repository

Star
 (10.602 stelle)
Metriche merge PR
 (Merge medio 4g 5h) (20 PR mergiate in 30 g)

Descrizione

Hi AutoGluon team,

I'm using TabularPredictor for temporally ordered tabular data and need more flexible cross-validation options. The current CV methods and "groups" parameter don't suffice for many time-based scenarios.

Standard k-fold CV can cause data leakage with ordered data. As far as I can tell, I'm unable to do the following using the "groups" parameter:

fold 1: training [month 1], test [month 2] fold 2: training [months 1-2], test [month 3] fold 3: training [months 1-3], test [month 4] ...and so on

Proposed solution: Add a custom_cv_strategy parameter to TabularPredictor, allowing custom splitting functions or pre-implemented strategies like forward chaining.

Let me know if I've overlooked something, and this is actually currently possible, or if there is a workaround.

Otherwise, thank you for considering!

Guida contributor