Interpretable machine learningOne major challenge for data science is to ensure that models do more good than harm. Along the way, we must ensure that models interpretations are valid, understand when they yield valid causal insights, and identify potential harms for subgroups. Our lab focuses on building tools for interpretable machine learning, which we view as a key component of trustworthy data science. These include powerful predictive models that are also interpretable, and improved methods to handle missing-not-at-random data and informative missing values. These methods have important implications in healthcare, where they have already produced novel and actionable clinical insights in cardiology. TalksSoftware
PapersThe Missing Indicator Method: From Low to High Dimensions Interpretable Survival Analysis for Heart Failure Risk Prediction Data-Efficient and Interpretable Tabular Anomaly Detection ControlBurn: Nonlinear Feature Selection with Sparse Tree Ensembles ControlBurn: Feature Selection by Sparse Forests Impact of Accuracy on Model Interpretations ``Why should you trust my explanation?'' Understanding uncertainty in LIME explanations Fairness Under Unawareness: Assessing Disparity When Protected Class Is Unobserved Causal Inference with Noisy and Missing Covariates via Matrix Factorization |