Classifier Chain Networks for Multi-Label Classification

In contrast to binary and multi-class classification, where each observation in the data is assigned to a single class, an observation in a multi-label classification task can have multiple labels. This type of problem arises in different fields, such as object detection in images, text analysis, bioinformatics, and recommendation systems (Tsoumakas et al., 2010). Consequently, numerous methods have been developed to handle multi-labeled outcomes. In contrast to existing methods, which often focus on modeling each outcome variable separately, our proposed method jointly models all labels to capture dependencies between them. In this study, we also refer to these dependencies between labels as label interdependencies. A frequently used method for a classification task with multi-labeled outcomes is to decompose the task into separate independent binary classifications (e.g., Boutell et al., 2004; Luaces et al., 2012). This approach is typically referred to as binary relevance. A limitation of binary relevance is the fact that it does not exploit potential correlations between the different labels (Godbole and Sarawagi, 2004; Zhang and Zhou, 2014).