SGLB: Stochastic Gradient Langevin Boosting

Authors: Aleksei Ustimenko, Liudmila Prokhorenkova

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We also empirically show that SGLB outperforms classic gradient boosting when applied to classification tasks with 0-1 loss function, which is known to be multimodal. and Our experiments on synthetic and real datasets show that SGLB outperforms standard SGB and can optimize globally such non-convex losses as 0-1 loss...
Researcher Affiliation Collaboration 1Yandex, Moscow, Russia 2Moscow Institute of Physics and Technology, Moscow, Russia 3HSE University, Moscow, Russia.
Pseudocode Yes Algorithm 1 SGB and Algorithm 2 SGLB
Open Source Code Yes The proposed algorithm is implemented within the Cat Boost open-source gradient boosting library (option langevin=True) (Cat Boost, 2020). and Our implementation of SGLB is available within the open-source Cat Boost gradient boosting library.
Open Datasets Yes The datasets are described in Table 1 of the supplementary materials.
Dataset Splits Yes We split each dataset into train, validation, and test sets in proportion 65/15/20.
Hardware Specification No The paper does not specify the hardware used for the experiments.
Software Dependencies No Our implementation of SGLB is available within the open-source Cat Boost gradient boosting library. (No version specified for CatBoost or other dependencies).
Experiment Setup Yes We set learning rate to 0.1 and ς = 10 1 for SLA. For SGLB, we set β = 103 and γ = 10 3. Moreover, we set the subsampling rate of SGB to 0.5. and For all algorithms, the maximal number of trees is set to 1000.