Abstract: Scoring systems are simple classification models that let users make quick predictions by adding and subtracting a few small numbers. These models are widely used in medicine and criminal justice, as experts find them easy to use and understand. In spite of extensive deployment, the vast majority of scoring systems are built using ad hoc approaches, which hinder predictive performance and fail to address critical constraints needed for adoption.
In this talk, I will describe two new machine learning methods to create scoring systems that perform as well as state-of-the-art predictive models, but are far easier to understand and customize:
- SLIM (Supersparse Linear Integer Models) to learn scoring systems for decision making.
- RiskSLIM (Risk-calibrated Supersparse Linear Integer Models) to learn scoring systems for risk assessment.
Unlike traditional machine learning methods, SLIM and RiskSLIM fit models by solving directly solving challenging discrete optimization problems. I will discuss how we can recover certifiably optimal solutions to these problems. I will then describe the benefits of this approach through several real-world applications, including recidivism prediction, ICU seizure prediction, and ADHD diagnosis.
Bio: Berk Ustun is a Postdoctoral Fellow at the Harvard University Center for Research in Computation for Society (CRCS). His research focuses on machine learning and causal inference in domains where humans have traditionally made decisions, such as healthcare, finance, and criminal justice. In particular, he is interested in designing new methods for algorithmic decision-making that let practitioners handle societal constraints, such as fairness, interpretability, and accountability. Berk holds a PhD in Electrical Engineering and Computer Science from MIT, an MS in Computation for Design and Optimization from MIT, and BS degrees in Operations Research and Economics from UC Berkeley.