0

In this video, I learned that drift monitoring would be easier in batch prediction than that in online prediction:
enter image description here

But I don't know why and I cannot find any information about it googling. In my opinion, in online prediction we only need to keep the prediction records and we have all the history data for drift analysis. What are the differences here between the two types of drift monitoring?

Lerner Zhang
  • 877
  • 1
  • 7
  • 19

1 Answers1

1

You're correct that, if you keep a record of all the predictions you've made, then you can monitor drift when doing online predictions. However, you probably need an extra database to log the predictions, and also a periodic job which calculates drift reading off that database. That can make for a complicated setup.

With batch predictions, if your batch is large enough, you can calculate summary statistics on that batch and detect drift at the same time as performing the prediction task. This is a lot easier to implement.

Raphael Lopez Kaufman
  • 600
  • 1
  • 9