Open choi328328 opened 1 year ago
Hi Jin Choi, I love the idea of adding in lightGBM!
Did you want to help add it in (no worries if not, but I wanted to see whether I could support you if you do)? To add a new classifier, we need a set function that specifies the hyper-parameters that users can search and seed, a fit function that takes the settings and data then fits the model and possibly a predict function that takes new data and the model to return the predictions. More details can be found here: https://ohdsi.github.io/PatientLevelPrediction/articles/AddingCustomModels.html
Of course. I will upload the code after developing it.
Commits have been requested to merge. The next steps are:
Hello, I'm Jin Choi from Ajou university.
Currently, the default gradient boosting machine algorithm is xgboost. I think we can use lightGBM as default, which was developed by Microsoft and cited more than 5000 times. (https://proceedings.neurips.cc/paper/2017/hash/6449f44a102fde848669bdd9eb6b76fa-Abstract.html)
LightGBM is known for 10x faster training times than xgboost while maintaining almost the same model performance. You can find comparisons between xgboost and lightGBM in the below link. (https://neptune.ai/blog/xgboost-vs-lightgbm)
Also, LightGBM has an official R package. (https://lightgbm.readthedocs.io/en/latest/R/index.html)
I think this change could accelerate model development with limited computational resources.