There was an interesting arxiv paper introducing sparrow which is another iteration of gradient boosted decision trees
https://arxiv.org/pdf/1901.09047.pdf
The premise is that this iteration can handle massive datasets using sampling and other techniques while using a small memory footprint. Food for thought as to whether the suggestions could be implemented in LightGBM.
I would like to take this task. Does anyone mind? Seems to be relevant to my interests
Hoping your offer of help gets sanctioned. This change is a potential gamechanger
Hi @GUEEN !
How is your process going?
Working on it. I tried to run the original rust code, but couldn't because of some error.
Now I am building a separate C++ sparrow library following the paper directly (please see my repository). After that, I'd like to integrate it into Lightgbm. Still a couple of weeks away, sorry.
Hi Gueen any further progress?
@MotoRZR Why did you close the issue?
Sorry didn鈥檛 realise I did. Must have been a fat finger on my part. Will reopen
Hello. Sorry, I have been busy on some other projects lately. I have written a standalone C++ sparrow library that seems to work, but still requires some polishing and bugfixes. Planning to return soon.
@GUEEN Good news, keep doing! We'll be happy to merge this into LightGBM eventually.
Closed in favor of being in #2302. We decided to keep all feature requests in one place.
Welcome to contribute this feature! Please re-open this issue (or post a comment if you are not a topic starter) if you are actively working on implementing this feature.
@MotoRZR
For everyone's information, the sparrow algorithm has been implemented in CatBoost
https://github.com/microsoft/LightGBM/issues/2302#issuecomment-543641394
Can you please provide some references/details?
@GUEEN How is your progress going?
Sorry, I've been busy with other commitments recently. Still planning to return sometime
Most helpful comment
Working on it. I tried to run the original rust code, but couldn't because of some error.
Now I am building a separate C++ sparrow library following the paper directly (please see my repository). After that, I'd like to integrate it into Lightgbm. Still a couple of weeks away, sorry.