Assessing human decisions in tabular machine learning model creation

Abstract

Tabular machine learning (ML) models are widely used in high-stakes domains, yet little attention has been paid to the risks introduced by subjective decisions during model development. This dissertation investigates how choices in missing data treatment and feature selection influence model outcomes and establishes a way to mitigate the risks.Using a risk framework defined as Risk = Frequency X Magnitude of Consequence, I first conducted a case study that varies missing data mechanisms, the amount of missing data, missing data treatment methods, and feature selection methods. Results highlighted that missing data treatment methods can substantially alter selected feature subsets and downstream model performance, with certain combinations reducing predictive performance to near-random levels.Then, a survey of 70 ML modelers assessed how frequently risky assumptions occur in practice, finding that over half of the participants made risky assumptions. The majority of participants had limited familiarity with the missing data mechanisms, which was a main source of risk. To mitigate these risks, I designed a prototype missing data tool that calculates the possible missing data mechanisms, identifies features with informative missingness, provides a summary of missing data, and gives explicit missing data treatment recommendations. A controlled user study showed that, on average, the proposed tool significantly reduced the likelihood of risky assumptions.