Drift at the feature level
While data drift looks at the whole input, feature drift zooms in on individual columns. A single upstream pipeline change, a unit swap, or a new null pattern can corrupt one feature and quietly distort predictions.
What to monitor per feature
- Summary statistics like mean, standard deviation, and quantiles.
- Missing value rate and the share of unexpected categories.
- Range violations where values fall outside historical bounds.
Pinpointing the culprit
Aggregate drift scores tell you something moved. Per feature monitoring tells you which feature, which usually points straight at the failing data source.
Practical guidance
- Prioritize monitoring the features with the highest importance to the model.
- Watch for schema changes such as a renamed or dropped column.
- A frozen feature that never changes can be as suspect as one that swings wildly.
Feature drift monitoring is often the fastest route from a model alert to the actual broken table.
Key idea
Monitor each feature with per column statistics and bounds so a broken upstream source is pinpointed quickly, focusing effort on the highest importance features.