Access OOB data and OOB error calculations of Random Forest #140

Activity

1. mdabros commented on Mar 31, 2021

   

   mdabros

   on Mar 31, 2021

   Owner

   More actions

   Hi @MichaelBenAssor,

   Originally I decided to leave this feature out of the RandomForest implementation to keep memory consumption low, and leave the option of getting an unbiased estimate of the model error to the CrossValidation classes. But it is quite a nice feature of the original RandomForest, so it is something that could be cool to have. So if I can find a good solution that is optional, and does not add memory usage or performance degradation when running without it, I can add it.

   I have a few other features I want to complete before working on this though, and with this being a spare time effort, it might be a while before I get to it :)

   Best regards
   Mads
2. MichaelBenAssor commented on Apr 1, 2021

   

   MichaelBenAssor

   on Apr 1, 2021

   Author

   More actions

   Hi @mdabros
   You can keep the OOB indices of records (record indices corresponding to the training set record lines) for each tree in a structure like:
   dictionary<tree#,List>
   This should keep the memory consumption to a minimum

   Thank you and I hope to see this wonderful project continue to flourish!

   Best,
   Michael

   👍React with 👍1Reacted by Mads Dabros