TreeGrad：将树集成迁移至神经网络
机器学习
2020-02-06 v3 机器学习
摘要
梯度提升决策树(GBDT)是流行的机器学习算法，有 LightGBM 等实现，并包含于 Scikit-Learn 等流行机器学习工具包中。许多实现只能以离线且贪婪的方式生成树。我们探索以最小性能损失将现有 GBDT 实现转换为已知神经网络架构的方法，以使决策分裂能够以在线方式更新，并提供扩展将分裂点改变作为神经架构搜索问题。我们给出了神经网络的泛化界。
引用
@article{arxiv.1904.11132,
title = {TreeGrad: Transferring Tree Ensembles to Neural Networks},
author = {Chapman Siu},
journal= {arXiv preprint arXiv:1904.11132},
year = {2020}
}
备注
Technical Report on Implementation of Deep Neural Decision Forests Algorithm. To accompany implementation here: https://github.com/chappers/TreeGrad. Update: Please cite as: Siu, C. (2019). "Transferring Tree Ensembles to Neural Networks". International Conference on Neural Information Processing. Springer, 2019. arXiv admin note: text overlap with arXiv:1909.11790