Multi-Objective Classification with Info-Fuzzy Networks

Abstract

The supervised learning algorithms assume that the training data has a fixed set of predicting attributes and a single-dimensional class which contains the class label of each training example. However, many real-world domains may contain several objectives each characterized by its own set of labels. Though one may induce a separate model for each objective, there are several reasons to prefer a shared multi-objective model over a collection of single-objective models. We present a novel, greedy algorithm, which builds a shared classification model in the form of an ordered (oblivious) decision tree called Multi-Objective Info-Fuzzy Network (M-IFN). We compare the M-IFN structure to Shared Binary Decision Diagrams and bloomy decision trees and study the information-theoretic properties of the proposed algorithm. These properties are further supported by the results of empirical experiments, where we evaluate M-IFN performance in terms of accuracy and readability on real-world multi-objective tasks from several domains.