True path rule hierarchical ensembles

Hierarchical classification problems gained increasing attention within the machine learning community, and several methods for hierarchically structured taxonomies have been recently proposed, with applications ranging from classification of web documents to bioinformatics. In this paper we propose a novel ensemble algorithm for multilabel, multi-path, tree-structured hierarchical classification problems based on the true path rule borrowed from the Gene Ontology. Local base classifiers, each specialized to recognize a single class of the hierarchy, exchange information between them to achieve a global “consensus” ensemble decision. A two-way asymmetric flow of information crosses the tree-structured ensemble: positive predictions for a node influence its ancestors, while negative predictions influence its offsprings. The resulting True Path Rule hierarchical ensemble is applied to the prediction of gene function in the yeast, using the FunCat taxonomy and biomolecular data obtained from high-throughput biotechnologies.