Benchmarking missing-values approaches for predictive models on health databases

RT @GaelVaroquaux: Reading bad empirical studies of missing-value imputation for predictive modeling, I want to point ours, which is a good…

Reading bad empirical studies of missing-value imputation for predictive modeling, I want to point ours, which is a good one 🤪 https://t.co/RFaBn1qXPG https://t.co/OJgXvo9bUk

@CatalystCoop @EIAgov @scikit_learn We often find that exposing missing values to the HistGradientBoosting yields better prediction than imputation https://t.co/RFaBn1rvFe

@caglar_ee @FrankRHutter Yup, I do believe that it does (so, IRL, the gap is wider than the one we report). We did a benchmark for missing values, and found that native handling of missing values in trees was the best performer: https://t.co/RFaBn1qXPG