Hybrid Binary-Chain Multi-label Classifiers
Pablo Hernández-Leal, Felipe Orihuela‐Espina, Luis Enrique Sucar, Eduardo F. Morales · University of Birmingham Research Portal (University of Birmingham) · 2012
In multi-label classification the goal is to assign an instance to a set of di↵erent classes. Several approaches have been proposed to deal with multi-label classification problems, ranging from considering each class independently from the other (binary relevance methods) to considering all the possible combinations of values of the original classes into a single compound class (power-set approach). In between, other methods have been proposed to consider dependencies among classes whilst trying to keep computational complexity of the method low. In this paper, instead of finding probabilistic dependencies among classes, we focused on finding independencies among classes using a simple correlation approach. We first build a correlation matrix among classes and use it to build chain classifiers among correlated sub-sets of classes while learning independent classifiers for uncorrelated classes. It is experimentally shown that this simple hybrid approach exhibits very competitive predictive performance among state-of-the-art multi-label classifiers with lower time complexity.