Data drift

How is drift calculated ?

Verta calculates drift using popular algorithms like Cosine distance and KL Divergence. Users have the ability to easily select (or modify) which algorithm to use from the monitoring dashboard. Drift is computed by measuring distribution changes between the model’s production values against a reference distribution and is supported for all features and output prediction columns.

The reference distribution can be your training, test or validation dataset and should be logged when you are registering a specific model version. Learn more about how to log reference data.

Visualizing and debugging drift

You can visualize the following drift charts for easier debugging and root cause analysis:

• A tabular view of all features and predictions with drift value

• A summary histogram comparing reference and production distribution

• Time series chart showing distribution over time (depending on the data type for example between binary vs float different types of visualization are supported)

• Line chart with drift value over time