Data Drift: Monitor Model Performance Changes with Fiddler's Insights
PreviousEnhance ML and LLM Insights with Custom MetricsNextEnsuring Data Integrity in ML Models And LLMs
Last updated
Was this helpful?
Last updated
Was this helpful?
Model performance can be poor if models trained on a specific dataset encounter different data in production. This is called data drift and it is a metric which is available on model inputs, outputs, and custom features. On the Insights dashboard for your model, Fiddler gives you a diverse set of visuals to explore different metrics.
Leverage the data drift chart to identify what data is drifting, when it’s drifting, and how it’s drifting. This is the first step in identifying possible model performance issues.
Fiddler supports the following:
Drift Metrics
Population Stability Index (PSI)
A drift metric based on the multinomial classification of a variable into bins or categories. The differences in each bin between the baseline and the time period of interest are then utilized to calculate it as follows:
🚧 Note
There is a possibility that PSI can shoot to infinity. To avoid this, PSI calculation in Fiddler is done such that each bin count is incremented with a base_count=1. Thus, there might be a slight difference in the PSI values obtained from manual calculations.
Average Values – The mean of a field (feature or prediction) over time. This can be thought of as an intuitive drift score.
Drift Analytics – You can drill down into the features responsible for the prediction drift using the table at the bottom.
Feature Impact: The contribution of a feature to the model’s predictions, averaged over the baseline. The contribution is calculated using random ablation feature impact.
Feature Drift: Drift of the feature, calculated using the drift metric of choice.
Prediction Drift Impact: A heuristic calculated using the product of the feature impact and the feature drift. The higher the score, the more this feature is likely to have contributed to the prediction drift.
In the Root Cause Analysis table of your drift charts, you can select a feature to see the feature distribution for both the time period under consideration and the baseline dataset.
Data drift is a great proxy metric for performance decline, especially if there is delay in getting labels for production events. (e.g. In a credit lending use case, an actual default may happen after months or years.)
Monitoring data drift also helps you stay informed about distributional shifts in the data for features of interest, which could have business implications even if there is no decline in model performance.
High drift can occur as a result of data integrity issues (bugs in the data pipeline), or as a result of an actual change in the distribution of data due to external factors (e.g. a dip in income due to COVID). The former is more in our control to solve directly. The latter may not be solvable directly, but can serve as an indicator that further investigation (and possible retraining) may be needed.
You can drill down deeper into the data by examining it in the Analyze tab.
