Building Machine Learning Pipelines

Chapter 13. Feedback Loops

Now that we have a smooth pipeline for putting a machine learning model into production, we don’t want to run it only once. Models shouldn’t be static once they are deployed. New data is collected, the data distribution changes (described in Chapter 4), models drift (discussed in Chapter 7), and most importantly, we would like our pipelines to continuously improve.

Adding feedback of some kind into the machine pipeline changes it into a life cycle, as shown in Figure 13-1. The predictions from the model lead to the collection of new data, which continuously improves the model.

Figure 13-1. Model feedback as part of ML pipelines

Without fresh data, the predictive power of a model may decrease as inputs change over time. The deployment of the ML model may in fact alter the training data that comes in because user experiences change; for example, in a video recommendation system, better recommendations from a model lead to different viewing choices from the user. Feedback loops can help us collect new data to refresh our models. They are particularly useful for models that are personalized, such as recommender systems or predictive text.

At this point, it is extremely important to have the rest of the pipeline set up robustly. Feeding in new data should cause the pipeline to fail only if the influx of new data causes the data statistics to fall outside ...