Target and Count encodings for categorical features - Iteration 1 #4964
Conversation
| const std::string count_prior_weight_key = "prior_weight"; | ||
|
|
||
| namespace LightGBM { | ||
| double CategoryFeatureTargetEncoder::Encode(double feature_value) { |
There was a problem hiding this comment.
Is the fold ID considered for encoding of training dataset?
There was a problem hiding this comment.
Ok. I see that the encoder is per-fold. So there's no question.
|
@tongwu-msft Thank you for working on this. The design overall looks good to me. Can we move on to integrating with the dataset classes? |
Yes, let me split the big PR into multiple iterations, this is it1 and we can do it parallel. |
tongwu-sh
requested review from
btrotta,
guolinke,
henry0312 and
huanzhang12
as code owners
tongwu-sh
requested review from
hzy46,
jameslamb,
Laurae2 and
StrikerRUS
as code owners
|
@tongwu-msft is this still WIP? |
|
I'm very confused. After 7 months of inactivity, this PR was closed without any comment...and it was supposed to replace #3234 which has not received a commit in about 8 months. @shiyu1994 @tongwu-msft @guolinke what is happening with the large changes to LightGBM's handling of categorical values? Is someone working on this and should it still be considered a requirement for v4.0.0 (#5153)? |
|
I can pick this up. But I'll first focus on the CUDA parts before v4.0.0. Another option would be extracting the categorical feature handling into a separate module for LightGBM dataset preprocessing (something like https://github.com/microsoft/LightGBM-transform). So that we can focus on the core training algorithm of GBDT in our codebase. |
|
This pull request has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this. |
Split PR: Target and Count encodings for categorical features to multiple part.
Implementation