Classificação de alto nível baseada em redes complexas para aprendizado multirrótulo

Abstract

Data classification is one of the most important topics in machine learning (ML) and aims to automate discrete learning tasks by assigning a class (or label) that characterizes each instance of the problem addressed. Traditional classification algorithms (or single-label) assume that each instance is associated with a single class, however, many real-world problems can be related to multiple labels simultaneously, such as the image annotation with multiple objects. As it is an extension of the single-label classification, most of the multi-label learning algorithms (MLL) are based on traditional classification techniques, inheriting their advantages but also their limitations. In relation to the limitations, most single-label classification techniques have a learning process guided only by physical characteristics of the data (e.g., distance or distribution) and ignore semantic and structural relationships of the data, such as pattern formation. Recently, several researches on ML have employed concepts of complex networks in order to capture structural and topological relationships of the data (i.e., high-level characteristics) and consequently improve their results. Inspired by the emerging usage of complex networks in ML, this dissertation investigates new methods based on complex networks for MLL, presenting new techniques for modeling the multi-label problem into a network as well as a new hybrid approach able to consider both physical and topological aspects of the data by combining complex networks with traditional MLL techniques. Experiments performed on artificial and real-world databases demonstrate the ability of the high-level technique to detect multiple patterns in the data and, as a result, improve the predictive performance of traditional MLL techniques. Moreover, this work paves a way to new developments based on complex networks to MLL.