A Graphical Model Approach for Multi-Label Classification

Abstract

Multi-Label (ML) classification problem is the assignment of many labels to a given sample from a fixed label set. It is considered as the more general version of the Multi-Class (MC) classification problem and its practical application areas vary from medical diagnosis to paper keyword selection. The general structure of an ML classification system involves transforming the problem into simpler MC and Single-Class (SC) problems. One such method is the Binary Relevance (BR) method that treats each label assignment as an independent SC problem, which makes BR systems scalable, but not accurate for some cases. This paper addresses the label independence problem of BR by assuming the outputs of each SC classifiers as observation nodes of a graphical model. The final label assignments are obtained by standard powerful Bayesian inference from the unobservable node. The proposed system was tested on standard ML classification datasets that produced encouraging results.