Summary

In this chapter, we explored the concepts of bias, explainability, fairness, and lineage. We started off by examining some of the common types of bias that can occur at various steps in the ML model development life cycle. This included sources of bias such as pre-existing bias, algorithmic bias, and collection or measurement bias, which further included sub-categories such as sampling bias, response bias, and observer bias. We talked about how to inspect for bias, using techniques such as data exploration and DIA.

Next, we dived into the use of explainability techniques to understand how our models make their decisions at inference time and to assess their fairness, particularly with regard to understanding how the input features in our dataset could influence our models’ predictions. We used tools such as PDPs and SHAP for these purposes. We then looked at how to use Vertex AI to get explanations from models that were hosted on Vertex AI endpoints. Going beyond...