Hyper-heuristics and decision trees for hierarchical multi-label classification problems

• Grant number: 16/14152-5
• Support Opportunities: Research Grants - Visiting Researcher Grant - International
• Start date: October 16, 2016
• End date: October 30, 2016
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Márcio Porto Basgalupp
• Grantee: Márcio Porto Basgalupp
• Visiting researcher: Leander Schietgat
• Visiting researcher institution: University of Leuven, Leuven (KU Leuven) , Belgium
• Host Institution: Instituto de Ciência e Tecnologia (ICT). Universidade Federal de São Paulo (UNIFESP). Campus São José dos Campos. São José dos Campos , SP, Brazil
• City of the host institution: São José dos Campos

Abstract

Hierarchical Multi-label Classification (HMC) is a complex problem, in which classes involved are structured in a hierarchy with hundreds or even thousands of classes. Additionally, instances can be simultaneously classified into more than one path in this hierarchy. These problems are very common, for example, in protein function prediction and annotation of medical images. Among the different algorithms that can be used in these problems, decision tree induction algorithms are a good choice, due their robustness and efficiency, and also because they produce interpretable models with satisfactory performances. However, there are still many open questions about the use of these algorithms in the HMC context, such as which stop and prune criteria to use, which split to use in an internal node, and how to consider the relationships between classes. In addition, only the top-down strategy was used until now. Given such many configuration possibilities, this project aims at implementing a hyper-heuristic for the construction of decision tree induction algorithms, tailored to HMC problems. In contrast to meta-heuristics, hyper-heuristics operate in a higher abstraction level, being used in the search for the best combination of components in the space of possibilities. These components are used to construct the decision tree induction algorithms. (AU)