Online Classification with Complex Metrics

TL;DR

This paper introduces a unified framework for online and batch binary classification optimizing complex, non-decomposable metrics like F-measure and Jaccard, with theoretical guarantees and practical efficiency.

Contribution

It extends thresholding characterization of optimal classifiers to complex metrics, providing gradient-based methods with finite-sample guarantees for both batch and online learning.

Findings

Convergence rate matches conditional probability estimation

Online algorithm achieves $O(1/\sqrt{n})$ sample complexity with logistic models

Empirical results outperform existing methods in accuracy and speed

Abstract

We present a framework and analysis of consistent binary classification for complex and non-decomposable performance metrics such as the F-measure and the Jaccard measure. The proposed framework is general, as it applies to both batch and online learning, and to both linear and non-linear models. Our work follows recent results showing that the Bayes optimal classifier for many complex metrics is given by a thresholding of the conditional probability of the positive class. This manuscript extends this thresholding characterization -- showing that the utility is strictly locally quasi-concave with respect to the threshold for a wide range of models and performance metrics. This, in turn, motivates simple normalized gradient ascent updates for threshold estimation. We present a finite-sample regret analysis for the resulting procedure. In particular, the risk for the batch case converges to the Bayes risk at the same rate as that of the underlying conditional probability estimation, and the risk of proposed online algorithm converges at a rate that depends on the conditional probability estimation risk. For instance, in the special case where the conditional probability model is logistic regression, our procedure achieves $O(\frac{1}{\sqrt{n}})$ sample complexity, both for batch and online training. Empirical evaluation shows that the proposed algorithms out-perform alternatives in practice, with comparable or better prediction performance and reduced run time for various metrics and datasets.