Wasserstein distance measures dissimilarity between probability distributions by estimating the minimum cost of transforming one distribution into another. In production ML, it compares a baseline reference distribution with a later current distribution to detect data drift. section 6.2
Use the distance as an operational drift signal:
- A larger distance indicates greater distributional drift; a smaller distance indicates greater similarity. section 6.2
- Evidently reports scores from 0 to infinity and uses 0.1 as its default drift-detection threshold. section 6.2.1
- Compare each feature between reference and inference data, then inspect the resulting score in a report or dashboard. section 6.2.1
Monitoring feature drift with Evidently
Create a DataDriftPreset report, run it with reference_data and current_data, and save the result as drift_report.html. The report compares every column and records the selected statistical test and drift score. section 6.2.1
