The model’s behavior is fluctuating. But is it necessarily the model’s fault? Regressions that appear to be rogue algorithms can ultimately be traced back to the source, the data.
So-called “data drift” occurs when production data no longer resembles the dataset used as training data. Perhaps the outliers that occur in the real world simply weren’t present during training.
(While “data drift” is a change in the distribution of the data itself, the related “concept drift” refers to a change in the relationship between the input data and the target variable. “Prediction drift” refers to a change in the model output. IBM provides another complete guide to drift detection.)
Ultimately, the optimal model monitoring strategy employs techniques that monitor data distribution as a key factor in model performance. One common method is the Kolmogorov-Smirnov test. This is useful for detecting whether the distribution of real-time production data differs significantly from the model’s training distribution.
Most managed ML platforms, such as Sagemaker and Evidently AI (an open source option), have tools to explore your data in a variety of ways. IBM’s wastonx.governance solution also has tools to detect drift in production data.
