Quantized Training of Gradient Boosting Decision Trees

This is the repository for experiments of paper Quantized Training of Gradient Boosting Decision Trees. The implementation is based on LightGBM.

Instructions to Run Experiments

To run the experiments, please:

1. Clone this repository, including all the submodules.
2. Download the datasets from https://pretrain.blob.core.windows.net/quantized-gbdt/dataset.zip (the .zip file is 38.29 GB).
3. Build the CLI executable files of LightGBM, LightGBM-master, catboost, and xgboost according to the CLI building instructions of these tools.
4. Generate the bash scripts using experiments/generate_script.py.
5. Run the generated bash script.
6. Parse the results into markdown table using experiments/parse_logs.py.

Training logs and sample outputs of experiments in the paper are provided in experiments/sample_outputs.