← All research spotlights

Research spotlight · Unit 5

Cynthia Rudin

Duke University

Research field
Interpretable machine learning, decision making in criminal justice and health care

Cynthia Rudin builds machine learning models that people can read in full, such as short lists of if-then rules and scoring systems, and studies where they can replace black boxes. Her 2019 paper in Nature Machine Intelligence argues that for high-stakes decisions about people the accepted trade-off between accuracy and readability is largely a myth, that explanations fitted to a black box after the fact can be misleading, and that no black box should be deployed where a readable model of the same accuracy exists. The unit presents her comparison of the proprietary COMPAS risk score with a three-rule model found by her group's CORELS algorithm as the case for readable models.

In the masterclass

Section 4, after random forests, where the readability of a decision tree is weighed against the accuracy of a black box.

The research is presented in Unit 5: Classifiers and Decision Trees, in a research spotlight box that explains the result at the level of the unit and links it to what you have just learned.

The paper

Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5), 206‑215. https://doi.org/10.1038/s42256-019-0048-x

The scientists named here are not contributors to the AI Masterclass, are not affiliated with it, and have not endorsed it. The program presents and explains their published research, with references to the original papers.