On a robust gradient boosting scheme based on aggregation functions insensitive to outliers

One new robust scheme for constructing gradient boosting algorithms is proposed. It is based on applying differentiable mean estimates insensitive or low-sensitive to outliers when constructing a robust empirical risk functional. This allows applying the iterative reweighting method to search for the next basic function and its weight. This gradient boosting procedure permits one to find the desired dependence based on data that contain a relatively large proportion of outliers.