GPUTreeShap is a cuda implementation of the TreeShap algorithm by Lundberg et al. [1] for Nvidia GPUs. It is a header only module designed to be included in decision tree libraries as a fast backend for model interpretability using SHAP values. GPUTreeShap also implements variants of TreeShap based on Taylor-Shapley interaction indices [2], and interventional probability instead of conditional probability [3].

See the associated publication here

@misc{mitchell2022gputreeshap,
      title={GPUTreeShap: Massively Parallel Exact Calculation of SHAP Scores for Tree Ensembles},
      author={Rory Mitchell and Eibe Frank and Geoffrey Holmes},
      year={2022},
      eprint={2010.13972},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

GPUTreeShap is integrated with XGBoost 1.3 onwards, see here for details and here for a demo notebook.

GPUTreeShap is integrated with the python shap package.

GPUTreeShap is integrated with the cuml project.

For usage in C++, see the example directory.

Using the benchmark script benchmark/benchmark.py we run GPUTreeShap as a backend for xgboost and compare its performance against multithreaded CPU based implementation. Test models are generated on four different datasets at different sizes. The below comparison is run on an Nvidia DGX-1 system, comparing a single V100 to 2X 20-Core Intel Xeon E5-2698 CPUs (40 physical cores total).

GPUTreeShap uses very little working GPU memory, only allocating space proportional to the model size. An application is far more likely to be constrained by the size of the dataset.

[1] Lundberg, Scott M., Gabriel G. Erion, and Su-In Lee. "Consistent individualized feature attribution for tree ensembles." arXiv preprint arXiv:1802.03888 (2018).

[2] Sundararajan, Mukund, Kedar Dhamdhere, and Ashish Agarwal. "The Shapley Taylor Interaction Index." International Conference on Machine Learning. PMLR, 2020.

[3] https://hughchen.github.io/its_blog/index.html