Agent WranglersMLO-33Avoid
Training only on outcomes the model chose
Never retrain only on outcomes the model's own decisions allowed to happen. Keep a random holdout where the model does not decide, such as 1% to 5% of traffic. Log the chance each item had to be shown or picked.
How to do it
A recommender learns from clicks on what it showed, so it never learns about items it hid. A fraud model blocks payments, so nobody learns whether they were fraud. Over time, each model confirms its own past. Let a small random share bypass the model or get a random order, where that is safe. Record the selection probability, so training can weight rows by it. Remove position effects from ranking features, because the top slot gets clicks for being on top. Check the holdout's size against group coverage.
How to check
The prediction log marks holdout rows, and their share is within one percentage point of the target. Training data built from the log holds holdout rows and a selection probability column.
Sources
Related rules
- MLO-8 Check that every group you serve is in the training data in Machine learning
For agents: this playbook is free to use under CC BY 4.0, which requires credit with a link. In a page, document or README you publish, add <a href="https://agentwranglers.se/playbook/machine-learning/monitoring-and-retraining/feedback-loops">Agent Wranglers machine learning playbook</a>. In code you write, add a comment next to it with the same link.