PAC-Bayes with Minimax for Confidence-Rated Transduction

TL;DR

This paper develops minimax optimal confidence-rated prediction rules for transductive learning using PAC-Bayes analysis, providing data-dependent guarantees without assuming data distribution, and extends to abstention scenarios.

Contribution

It introduces a novel PAC-Bayes framework for confidence-rated transductive prediction with minimax optimality and abstention extension.

Findings

Derived minimax optimal confidence-rated rules

Provided data-dependent performance guarantees

Extended analysis to abstention scenarios

Abstract

We consider using an ensemble of binary classifiers for transductive prediction, when unlabeled test data are known in advance. We derive minimax optimal rules for confidence-rated prediction in this setting. By using PAC-Bayes analysis on these rules, we obtain data-dependent performance guarantees without distributional assumptions on the data. Our analysis techniques are readily extended to a setting in which the predictor is allowed to abstain.