Data science with statistical learning

Some of the research at Soda is on the statistical foundations of machine learning, in particular for problem important to data science. The goals here are to use machine-learning models as non-parametric estimators for common problems in data science. Beyond mere prediction accuracy, questions of statistical control arise.

Statistical learning with missing values

Statistical inference with missing values has been studied for decades, but modern machine-learning practice brings new trade-offs. In particular, we have shown that the classical view on imputation may not give the best-performing predictors, and that missing-not-at-random settings could be tackled by machine learning models.

Machine learning for causal inference

Modern causal inference builds on estimating response function, for treated and non treated individuals, or probability of treatment or trial inclusion. We study the use of machine-learning models to estimate these quantities. Indeed, as we deal with increasingly complex data, such as that in Electronic Health Records, simple parametric models are no longer enough to leverage the data: the data is made of multiple tables, with many missing values and non-normalized text inputs.

One specific problem that we have focused on is that of generalizing a effect inferred on a study sample with a selection bias compared to the target population. This question is related to external validity of a study.

Publications

Missing values

Causal inference