Naive Bayes Classification for Subset Selection in a Multi-label Setting

This article introduces a novel probabilistic formulation of multi-label classification based on the Bayes theorem. Under the naive hypothesis of conditional independence of features given the labels, a pseudo-bayesian inference approach is adopted, known as Naive Bayes. The prediction consists of two steps: the estimation of the size of the target label set and the selection of the elements of this set. This approach is implemented in the NaiBX algorithm, an extension of naive Bayes into the multi-label domain. Its properties are discussed and evaluated on real-world data.