Automatic Machine Learning for Multi-Label Classification

• Grant number: 24/19234-6
• Support Opportunities: Regular Research Grants
• Start date: February 01, 2025
• End date: January 31, 2027
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Mobility Program: SPRINT - Projetos de pesquisa - Mobilidade
• Principal Investigator: Ricardo Cerri
• Grantee: Ricardo Cerri
• Principal researcher abroad: Isaac Triguero Velazquez
• Institution abroad: University of Nottingham, University Park, England
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Associated researchers: Diego Furtado Silva
• Associated research grant: 22/02981-8 - Novelty detection in multi-label data streams classification, AP.PNGP.PI

Abstract

Automatic Machine Learning (AutoML) is an emerging research field that aims to find automatic Machine Learning (ML) solutions for various problems. Using limited computational resources, AutoML automates the construction of ML pipelines, including data preparation, feature engineering, model generation, and model evaluation tasks. AutoML can potentially reduce human errors when designing ML solutions, increasing confidence in the results and reassuring the audience. A challenging task that can benefit from AutoML is multi-label classification, where an instance can be simultaneously classified into two or more classes. Various applications can be found in bioinformatics, image and text processing, and music and product categorization. Although different strategies exist to find the best classifiers for multi-label problems, these are usually specialized in specific datasets and use a limited search space of classifiers. In this project, we will use AutoML as a General Purpose Artificial Intelligence System (GPAIS), proposing a new method that acts like an Artificial Intelligence expert, determining which algorithms and/or components (and hyperparameters) are most suitable for a set of multi-label classification datasets for different domains. This is known as closed-world GPAIS. Our proposal will use different strategies, such as multi-objective optimization to search the space of possible solutions and surrogate models to speed up execution times. This is the first proposal for a general purpose AutoML system tailored to domains containing different multi-label classification datasets. The Brazilian partners of this proposal are researchers in the IARA Project (Artificial Intelligence in the Remaking of Urban Environments) funded by FAPESP, a joint nationwide collaboration involving universities and companies in Brazil. Thus, the classifiers constructed can be applied to smart city-related problems, such as categorizing construction materials left on illegal locations, such as sidewalks. (AU)