← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 7011.i · 生成模型:DeepSeek-V4.1-Flash

可解释性与监控的代价

理解等级:高级

最重要的服务前建议是:业务 KPI 应与模型指标相关,且能从开发连续追踪到生产——只有基础设施洞察而无业务洞察的模型近乎无用,只在开发里表现好却看不清生产行为的模型 arguably 比无用更有害;做数据完整性监控时,最重要的特征应被赋予高优先级(如打车 ETA 模型里接客地点要在整个周期内清晰可得)。可解释性对监控另有一层难题:完整可解释性在生产中比开发更耗资源、更慢更贵,而人往往又最急迫地需要它,用途包括确定哪些特征该优先保障数据完整性、调查某条或某切片预测、以及回应业务方对解释性的要求。因代价高且业务方未必具备背景,有效做法常是与他们谈清真正关心什么,很多情况下无需深入建模细节(甚至可专门建监控方案)就能回应。但排障时解释性不可替代:借贷申请可能因「申请日期=2 月 29 日」这一罕见取值被拒——该特征通常不在全局前十重要特征里,可若训练集里 2 月 29 日的少量申请后来都是坏风险,模型就可能重度依赖它拒贷。可借定期记录的预测摘要、LIME/SHAP 等代理/重复模型或定制方案暴露驱动因素,风险与合规团队也用得上,能帮助非技术用户理解系统性问题,可惜多数在生产中太贵。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • Monitoring and Observability for Models · 查看原书依据

      People responsible for ML models want explainability for a few reasons: establishing which features should be prioritized for data integrity, investigating a specific prediction or specific slices of predictions, and responding to business requirements for explainability generally. Since it’s so expensive, and since business people may not have all the background required to understand it, being effective here often amounts to

    参考来源

    - 位置:《Reliable Machine Learning》第09章《Monitoring and Observability for Models》(源文件:extracted/Reliable Machine Learning/_epub-src/OEBPS/ch09.xhtml) - 结论依据:原文明确阐述该主题,并给出定义、实践或反例。 - 原始内容:People responsible for ML models want explainability for a few reasons: establishing which features should be prioritized for data integrity, investigating a specific prediction or specific slices of predictions, and responding to business requirements for explainability generally. Since it’s so expensive, and since business people may not have all the background required to understand it, being effective here often amounts to