Expressive Power of Binary Relevance and Chain Classifiers Based on Bayesian Networks for Multi-label Classification

Abstract

Abstract. Bayesian network classifiers are widely used in machine learning because they intuitively represent causal relations. Multi-label classification problems require each instance to be assigned a subset of a defined set of h labels. This problem is equivalent to finding a multivalued decision function that predicts a vector of h binary classes. In this paper we obtain the decision boundaries of two widely used Bayesian network approaches for building multi-label classifiers: Multi-label Bayesian network classifiers built using the binary relevance method and Bayesian network chain classifiers. We extend our previous single-label results to multi-label chain classifiers, and we prove that, as expected, chain classifiers provide a more expressive model than the binary relevance method.

Language Models Are An Effective Patient Representation Learning Technique For Electronic Health Record Data

“…There is a large body of literature on multi-label problems and a number of techniques for modeling them. For the purposes of our experiments, we choose the simplest possible technique of transforming this problem into a binary classification problem through the binary relevance method [33]. We model p(…”

Decision functions for chain classifiers based on Bayesian networks for multi-label classification

“…Moreover, we suggest some theoretical reasons why the simple binary relevance method can perform poorly when relationships among labels exist, and we prove that chain classifiers provide more expressive models. A broader chain classifiers class than in Varando et al [21] is considered and studied extensively and a bounding on the expressive power of those models is proved. Moreover we present novel illustrative examples both about the one-dimensional results and about multi-label ones.…”

Theoretical studies on Bayesian network classifiers

“…• G. Varando Chapter 3 is directly derived with few changes from . Chapter 4 includes mainly the content of Varando et al [2016] and Varando et al [2014] plus an additional section containing a result published in Borchani et al [2015]. Chapter 5 contains the work that is currently under preparation as Varando et al 2018.…”