forust
forust copied to clipboard
jinlow
A lightweight gradient boosted decision tree package.
Hi there! I'm using the forust library and have noticed a curious behavior regarding memory consumption. Unlike popular libraries like XGBoost and LightGBM, where memory usage increases significantly with a...
I guess the following implementation has O(n) complexity. Correct me if I am wrong. It passes all the tests for `pivot_on_split`. It became hacky to handle edge cases. If you...
For some reason, Shapley runs much slower on windows than Linux. ```python import forust import xgboost as xgb import seaborn as sns import numpy as np df = sns.load_dataset("titanic") X...
Would you be open to adding a feature allowing for the option of serializing a GradientBooster struct using MessagePack instead of JSON? If so, I'd be happy to submit a...
Consider accepting `eval_set` and `sample_weight_eval_set` in the same way XGBoost scikit-learn does. This will just make the API easier to use directly in XGBoost scikit-learn pipelines.
This paper could be useful for implementing an unbiased feature importance measure https://arxiv.org/pdf/2305.10696.pdf
Following the suggestion in this blog post: https://www.statworx.com/en/content-hub/blog/how-to-speed-up-gradient-boosting-by-a-factor-of-two/
This parameter would allow users to specify the range over which a variable can be split. This way splits would never be made at large values that may not be...
To support different evaluation strategies, could a trait be implemented that abstracts and captures the evaluation flow logic to support other paradigms.
Currently in the gradientbooser fit method, all of the data is used for determining cuts for binning the data. It would like speed things up, if we allowed for a...