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.