Learning with risks based on M-location
This work provides a new, interpretable, and adjustable risk function class for machine learning practitioners seeking more robust and flexible risk management than traditional mean-variance approaches.
This paper introduces a new class of risk functions that generalize beyond mean-variance, based on the location and deviation of the loss distribution. This class is easily implemented, offers finite-sample stationarity guarantees for stochastic gradient methods, and significantly impacts the test loss distribution.
In this work, we study a new class of risks defined in terms of the location and deviation of the loss distribution, generalizing far beyond classical mean-variance risk functions. The class is easily implemented as a wrapper around any smooth loss, it admits finite-sample stationarity guarantees for stochastic gradient methods, it is straightforward to interpret and adjust, with close links to M-estimators of the loss location, and has a salient effect on the test loss distribution.