Skip to yearly menu bar Skip to main content


Poster

Regret Bounds for Multilabel Classification in Sparse Label Regimes

RĂ³bert Busa-Fekete · Heejin Choi · Krzysztof Dembczynski · Claudio Gentile · Henry Reeve · Balazs Szorenyi

Hall J (level 1) #810

Keywords: [ sparse multilabel classification ] [ regret bounds ]


Abstract: Multi-label classification (MLC) has wide practical importance, but the theoretical understanding of its statistical properties is still limited. As an attempt to fill this gap, we thoroughly study upper and lower regret bounds for two canonical MLC performance measures, Hamming loss and Precision@$\kappa$. We consider two different statistical and algorithmic settings, a non-parametric setting tackled by plug-in classifiers \`a la $k$-nearest neighbors, and a parametric one tackled by empirical risk minimization operating on surrogate loss functions. For both, we analyze the interplay between a natural MLC variant of the low noise assumption, widely studied in binary classification, and the label sparsity, the latter being a natural property of large-scale MLC problems. We show that those conditions are crucial in improving the bounds, but the way they are tangled is not obvious, and also different across the two settings.

Chat is not available.