Improving public dataset characterization using OpenML

• Grant number: 23/11801-6
• Support Opportunities: Scholarships abroad - Research Internship - Scientific Initiation
• Start date: December 01, 2023
• End date: February 29, 2024
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal Investigator: Ana Carolina Lorena
• Grantee: Nathan Falcão Carvalho
• Supervisor: Joaquin Vanschoren
• Host Institution: Divisão de Ciência da Computação (IEC). Instituto Tecnológico de Aeronáutica (ITA). São José dos Campos , SP, Brazil
• Institution abroad: Eindhoven University of Technology (TU/e), Netherlands
• Associated to the scholarship: 23/03958-2 - Gathering meta-data from public repositories, BP.IC

Abstract

In the past few years, Machine Learning (ML) has become a popular solution for problems in an extremely diverse range of fields, such as Biomedical Informatics, Natural Language Processing, and website recommendation systems. An essential step in ML is determining which algorithm best suits each specific domain, along with what hyperparameters maximize the desired result. The field of Meta-Learning (MtL) solves this algorithm selection problem by analyzing several algorithmic performance values on a diverse range of datasets. Learning from this metadata produces helpful insights into the process of selecting an algorithm to solve a specific problem. For a MtL study to be done, it is necessary to assemble metadata comprising of algorithmic performances and dataset characteristics. Since locally running a big range of ML algorithms on several datasets can be computationally demanding, a common way to acquire such data is to use public dataset repositories, of which OpenML is a popular option. Nevertheless, there are still limitations to this approach, as up-to-date dataset characteristics, called meta-features, can often be hard to get. Since several tools for extracting more useful and representative meta-features from datasets are already documented in the literature, we plan to enrich the OpenML repository by diversifying the range of available metadata. (AU)