[论文] Quantifying structural stability of transparent models for general ins...
研究领域: ML 作者: Félix Côté, Marie-Pier Côté 发布时间: 2025-11-13 arXiv: 2511.10777
论文概要
研究领域: ML 作者: Félix Côté, Marie-Pier Côté 发布时间: 2025-11-13 arXiv: 2511.10777
中文摘要
一般保险合同期限短,导致数据分布快速变化,需要定期更新模型。我们引入函数透明性的概念,并提出均方形状函数距离来量化透明模型的结构稳定性。使用三个公开车险数据集,我们表明:带中心协变量的广义线性模型是稳定的;可解释提升机(EBM)在稳定性与性能间提供折衷;顺序建模与固定超参数等稳定化策略可显著改善 EBM 和 XGBoost 的稳定性,且对性能影响可忽略。
原文摘要
The short duration of general insurance contracts leads to rapid distributional shifts, warranting regular model updates. We introduce the notion of functional transparency and propose the Mean Squared Shape function distance to quantify structural stability of transparent models. Using three public auto insurance datasets, we show GLMs with centred covariates are stable, EBMs offer a stability-performance compromise, and stabilization strategies like sequential modelling and fixed hyperparameters substantially improve stability for both EBM and XGBoost with negligible performance impact.
*自动采集于 2026-11-15*
#论文 #arXiv #ML #小凯