Streaming Gradient Boosted Trees (SGBT)

• Paper: Gradient boosted trees for evolving data streams

News

• SGBT is also availabe on CapyMOA, Python stream learning framework.
• SGBT is now available on MOA main branch, Java stream learning framework.
• SGBR: SGBT with data stream regression support [paper] [code]

Run MOA experiments (Based on this fork of MOA)

Requirements

• Conda

Datasets

• datasets from Streaming Random Patches (SRP): https://github.com/hmgomes/StreamingRandomPatches/tree/master/datasets
• synthetic data sets:

LED_a_S="-s (ConceptDriftStream -s (generators.LEDGeneratorDrift -d 1)   -d (ConceptDriftStream -s (generators.LEDGeneratorDrift -d 3) -d (ConceptDriftStream -s (generators.LEDGeneratorDrift -d 5)  -d (generators.LEDGeneratorDrift -d 7) -w 50 -p 250000 ) -w 50 -p 250000 ) -w 50 -p 250000 -r $random_seed )"

LED_g_S="-s (ConceptDriftStream -s (generators.LEDGeneratorDrift -d 1)   -d (ConceptDriftStream -s (generators.LEDGeneratorDrift -d 3) -d (ConceptDriftStream -s (generators.LEDGeneratorDrift -d 5)  -d (generators.LEDGeneratorDrift -d 7) -w 50000 -p 250000 ) -w 50000 -p 250000 ) -w 50000 -p 250000 -r $random_seed )"

AGR_a_S="-s (ConceptDriftStream -s (generators.AgrawalGenerator -f 1) -d (ConceptDriftStream -s (generators.AgrawalGenerator -f 2) -d (ConceptDriftStream -s (generators.AgrawalGenerator )   -d (generators.AgrawalGenerator -f 4) -w 50 -p 250000 ) -w 50 -p 250000 ) -w 50 -p 250000 -r $random_seed )"

AGR_g_S="-s (ConceptDriftStream -s (generators.AgrawalGenerator -f 1) -d (ConceptDriftStream -s (generators.AgrawalGenerator -f 2) -d (ConceptDriftStream -s (generators.AgrawalGenerator )   -d (generators.AgrawalGenerator -f 4) -w 50000 -p 250000 ) -w 50000 -p 250000 ) -w 50000 -p 250000 -r $random_seed )"

RBF_m_S="-s (generators.RandomRBFGeneratorDrift -c 5 -s .0001 -r $random_seed -i $random_seed)"

RBF_f_S="-s (generators.RandomRBFGeneratorDrift -c 5 -s .001 -r $random_seed -i $random_seed)"

RBF_Bm_S="-s (generators.RandomRBFGeneratorDrift -c 2 -s .0001 -r $random_seed -i $random_seed)"

RBF_Bf_S="-s (generators.RandomRBFGeneratorDrift -c 2 -s .001 -r $random_seed -i $random_seed)"

RandomTreeGenerator_S="-s (generators.RandomTreeGenerator -r $random_seed -i $random_seed)"

RandomRBF5_S="-s (generators.RandomRBFGenerator -r $random_seed -i $random_seed -c 5)"

LED_S="-s (generators.LEDGenerator -i $random_seed)"

How to set up environment

From source root run:

bash ./moa/src/main/scripts/reinit_conda.sh <conda_env_path> ./moa/src/main/scripts/conda.yml

How to build MOA

From source root run:

bash ./moa/src/main/scripts/build_moa.sh <maven_repo_path> <conda_env_path>

Run GUI

bash moa/src/main/scripts/moa_gui_with_NN_support.sh <maven_repo_path> <conda_env_path> <djl_cache_dir>

Run experiments

From < results dir > run:

bash <moa_source_root>/moa/src/main/scripts/run_moa.sh <dataset_dir> <results_dir> <djl_cache_dir> <maven_repo_path> <conda_env_path>

Notes:-

<moa_source_root>/moa/src/main/scripts/run_moa.sh could be copied to any place and run.

Change dataset variable in run_moa.sh to change the data set.

Cite this work

@article{gunasekara2024gradient,
  title={Gradient boosted trees for evolving data streams},
  author={Gunasekara, Nuwan and Pfahringer, Bernhard and Gomes, Heitor and Bifet, Albert},
  journal={Machine Learning},
  volume={113},
  number={5},
  pages={3325--3352},
  year={2024},
  doi={https://doi.org/10.1007/s10994-024-06517-y}
  publisher={Springer}
}

Change learners variable in run_moa.sh to change learner command.