Learning Label Dependency and Label Preference Relations in Graded Multi-label Classification

Abstract

Graded multi-label classification (GMLC) is a supervised machine learning task where the association between each data and a label has a membership degree from an ordinal scale of membership degrees: for example, an odorous molecule can be associated to the graded subset of odors {strong musc, moderate animal} based on the ordinal scale of odor intensity: {very weak, weak, moderate, strong, very strong}, and a movie can be associated to the graded subset of labels {action $\star \star \star \star$, suspense $\star \star$, humour $\star \star$} based on the ordinal scale of one-to-five star rating. The aim in GMLC is to build a predictive model called classifier, in order to predict the graded set of labels based on descriptive attributes of data. For example, predicting the graded set of molecule odors based on molecular properties such as the molecular structure and weight. Or predicting the graded set of genres for a movie based on the synopsis and the main actors. An interesting challenge in GMLC is learning label relations and exploiting them to enhance the prediction performance of classifiers. A label relation can be a dependency relation: for example, movies containing a lot of ‘action’ often contains also some ‘suspense’. Another type of label relations is preference relations: for example, it is preferred to associate a movie containing a lot of movements to the label ‘action’ than to the label ‘humour’. The limitation of existing approaches is that they can either learn dependency relations or preference relations. This work reviews state of the art GMLC approaches, and introduces a new GMLC approach that can learn both dependency and preference label relations. Experiments on real datasets show that the new approach outperforms baseline approaches according the used prediction evaluation measures.