Summary

Data and model drift refer to a phenomenon that occurs when the statistical properties of a dataset or underlying model change over time. In this chapter, we reviewed how this can have an adverse impact on the predictions of models and, hence, on business outcomes. To make sure models function as desired, companies implement an ML life cycle that ensures design, development, deployment, and monitoring best practices are in place. Drifts can happen for a variety of reasons, including changes in the underlying population and changes in the way data is collected. When data drift happens, it can create bias in ML models that are trained on this data, which can be quite problematic for regulations and compliance.

In this chapter, we reviewed several ways to detect and mitigate bias due to data or model drift, and to monitor your training and validation error rates closely using different tools, including open source and commercial hyperscaler products. There are various other...