When data pipelines feed into machine learning models, data issues can occur that remain silent and undetected. Your machine learning model will never email you to alert you that “the data looks wrong”; it will just keep operating on data that isn’t up-to-date with the real world. Since the relationships from which machine learning models learn inevitably change over time, not noticing those issues generally cause models to decay.
In this presentation we will show you how you can detect and prevent model drift and ensure that your machine learning models don’t start degrading and failing once you put them into production, using a tried-and-tested framework to get ahead of silent data issues, before they have a major downstream impact on the business.
We’ll examine the common types of model drift and participants will learn how to detect, analyze, and manage model drift to ensure that their machine learning models are running reliably and efficiently in production. to deliver trusted data to the business using a (built or bought) data monitoring platform.
Bastien Boutonnet is a neuroscientist by training who turned full-time data science and developer tooling nerd for the last 6 years. At Soda.io he leads product domains around automated monitoring and data incidents resolution powered by machine learning and other if/elses machines. When he's not working he likes to think about design and architecture, music production, and streaming his open-source contributions on Twitch.