A template for launching Machine Learning into human decision making is to take all previous decisions as labeled data, featurize the space, and train a model. This template has created a 1:1 paradigm mapping between models and their purpose in human decision making for industries from aviation to law enforcement. Unfortunately, this paradigm is expensive and hard to scale. Only a subset of human decisions can meet the criteria of enough labeled data, a high priority enough use case to warrant the ML team building the model, and the budget for model infrastructure.
However, if we leverage parameterization of models and provide context for the model’s predictions, we can create a self-serve environment for people making a variety of types of decisions. This cracks the 1:1 paradigm and enables dozens of types of decisions to be supported in unison for applications like supply chains. We’ll explore best practices and use Stitch Fix’s Style Explorer as a case study.