Признаки в обучении и проде расходятся: воспроизводим ошибку на тестовом наборе

habrahabr.ru · science
Full story

A model can pass its offline tests and still fail in production when the input features it sees at prediction time drift from what it was trained on. The article walks through this problem, known as Training–Serving Skew, using Python as the working example. It covers how to detect the skew, rebuild the data as it looked at the moment of prediction, and automate checks before a release. A hands-on example is offered for readers to work through.

Модель проходит офлайн-тесты, но ошибается в продакшене из-за расхождения входных признаков. На примере Python разбираем, как выявить Training–Serving Skew, восстановить данные на момент прогноза и автоматизировать проверку перед релизом. Разобрать пример

Read full article → Compare