A MOM-based ensemble method for robustness, subsampling and hyperparameter tuning

Hyperparameter tuning and model selection are important steps in machine learning. Unfortunately, classical hyperparameter calibration procedures sensitive to outliers heavy-tailed data. In this work, we construct a procedure which can be seen as robust alternative cross-validation is based on median-of-means principle. Using procedure, also build an ensemble method which, trained with algorithms corrupted data, selects algorithm, trains it large uncorrupted subsample automatically tunes its hyperparameters. particular, the approach transform any into data while The construction relies divide-and-conquer methodology, making easily scalable even dataset. This tested LASSO known highly outliers.