Prediction with expert advice is a classic online learning framework where a learner aggregates predictions from K experts over T rounds. This paper considers a setting in which the losses are heavy-tailed, meaning they have a bounded (1+ε)-moment for some ε in (0,1], rather than a bounded variance or exponential moment.
The assumption is that the (1+ε)-moment is bounded by θ > 0. This is a weaker condition than many standard assumptions, making the results potentially applicable to data with extreme values. The abstract notes that existing algorithms are designed for a different regime, but the source truncates before explaining what that regime is.
The significance of this work lies in its theoretical framing of a common practical challenge: heavy-tailed losses appear in many online learning problems, and standard algorithms often rely on stronger tail assumptions. However, because the abstract is incomplete, the specific contributions or bounds are not available in the source.