An analysis of chaining in multi-label classification

The idea of classifier chains has recently been introduced as a promising technique for multi-label classification. However, despite being intuitively appealing and showing strong performance in empirical studies, still very little is known about the main principles underlying this type of method. In this paper, we provide a detailed probabilistic analysis of classifier chains from a risk minimization perspective, thereby helping to gain a better understanding of this approach. As a main result, we clarify that the original chaining method seeks to approximate the joint mode of the conditional distribution of label vectors in a greedy manner. As a result of a theoretical regret analysis, we conclude that this approach can perform quite poorly in terms of subset 0/1 loss. Therefore, we present an enhanced inference procedure for which the worst-case regret can be upper-bounded far more tightly. In addition, we show that a probabilistic variant of chaining, which can be utilized for any loss function, becomes tractable by using Monte Carlo sampling. Finally, we present experimental results confirming the validity of our theoretical findings