Approximation trees: statistical reproducibility in model distillation
Approximation trees: statistical reproducibility in model distillation
复制标题
DOI：
10.1007/s10618-022-00907-3
复制
发表时间：
2023-01
影响因子：
4.8
通讯作者：
Yichen Zhou;Zhengze Zhou;G. Hooker
中科院分区：
文献类型：
--
作者：
Yichen Zhou;Zhengze Zhou;G. Hooker
本文探讨了可重复性的学习解释黑箱预测通过模型蒸馏使用分类树。我们发现，常见的树蒸馏方法无法重现一个单一的稳定的解释时，适用于同一教师模型，由于蒸馏过程的随机性。我们研究这个问题的可靠解释，并提出了一个标准化的框架树蒸馏，以实现再现性。所提出的框架包括（1）稳定树分裂的统计测试，以及（2）当使用提供其预测的不确定性估计的教师时，用于树构建的停止规则，例如随机森林。我们证明了所提出的蒸馏方法在各种合成和真实世界的数据集上的经验性能。
This paper examines the reproducibility of learned explanations for black-box predictions via model distillation using classification trees. We find that common tree distillation methods fail to reproduce a single stable explanation when applied to the same teacher model due the randomness of the distillation process. We study this issue of reliable interpretation and propose a standardized framework for tree distillation to achieve reproducibility. The proposed framework consists of (1) a statistical test to stabilize tree splits, and (2) a stopping rule for tree building when using a teacher that provides an estimate of the uncertainty of its predictions, e.g. random forests. We demonstrate the empirical performance of the proposed distillation method on a variety of synthetic and real-world datasets.