Machine Learning Production Systems

Chapter 8. Model Analysis

Successfully training a model and getting it to converge feels good. It often feels like you’re done, and if you’re training it for a class project or a paper that you’re writing, you kind of are done. But for production ML, after the training is finished you need to enter a new phase of your development that involves a much deeper level of analysis of your model’s performance, from a few different directions. That’s what this chapter is about.

Analyzing Model Performance

After training and/or deployment, you might notice a decay in the performance of your model. In addition to determining how to improve your model’s performance, you’ll need to anticipate changes in your data that you might expect to see in the future, which are generally very domain dependent, and react to the changes that occurred since you originally trained your model.

Both of these tasks require analyzing the performance of your model. In this section, we’ll review some basics of model analysis. When conducting model analysis, you’ll want to look at model performance not just on your entire dataset, but also on smaller chunks of data that are “sliced” by interesting features. Looking at slices gives you a much better understanding of the variance of individual predictions than what you’d get by looking at your entire dataset.

Choosing the slices that are important to analyze is usually based on domain knowledge. Though slicing on any feature used by your model can provide insights, ...