A flexible class of dependence-aware multi-label loss functions

Abstract The idea to exploit label dependencies for better prediction is at the core of methods multi-label classification (MLC), and performance improvements are normally explained in this way. Surprisingly, however, there no established methodology that allows analyze dependence-awareness MLC algorithms. With goal mind, we introduce a class loss functions able capture important aspect dependence. To end, leverage mathematical framework non-additive measures integrals. Roughly speaking, measure modeling importance correct predictions subsets (instead single labels), thereby their impact on overall evaluation, flexible well-known Hamming subset 0/1 losses rather extreme special cases function class, which give full sets or entire set, respectively. We present concrete instantiations appear be especially appealing from perspective. assessment classifiers terms these illustrated an empirical study, clearly showing aptness capturing dependencies. Finally, while not being main also show some preliminary results minimization parametrized family losses.