Conformal on-line mannequin aggregation


arXiv:2403.15527v1 Announce Sort: new
Summary: Conformal prediction equips machine studying fashions with an inexpensive notion of uncertainty quantification with out making robust distributional assumptions. It wraps round any black-box prediction mannequin and converts level predictions into set predictions which have a predefined marginal protection assure. Nonetheless, conformal prediction solely works if we repair the underlying machine studying mannequin upfront. A comparatively unaddressed situation in conformal prediction is that of mannequin choice and/or aggregation: for a given downside, which of the plethora of prediction strategies (random forests, neural nets, regularized linear fashions, and so forth.) ought to we conformalize? This paper proposes a brand new method in direction of conformal mannequin aggregation in on-line settings that’s primarily based on combining the prediction units from a number of algorithms by voting, the place weights on the fashions are tailored over time primarily based on previous efficiency.

Supply hyperlink


Please enter your comment!
Please enter your name here