Understanding Target Variable Selection in Machine Learning Model Development

Predictive modeling has recently attracted a lot of attention from organizations trying to leverage AI and big data to improve their work processes such as decision-making. However, real-world problems are rarely well-formulated machine learning problems. Practitioners have to supply a well-defined predictive target to operationalize a predictive model. In such cases, they often resort to using observed variables to approximate the actual construct of interest. For example, people have used high sales numbers as a proxy for a good employee. Proxy label selection is a recurring challenge when predictive ML is applied to real-world problems. The purpose of this interview study is to understand how ML practitioners select proxy labels, evaluate proxy labels, and iterate through the different tasks involved.