Integrating an algorithm recommendation module into the PyISpace package

• Grant number: 24/16535-5
• Support Opportunities: Scholarships abroad - Research Internship - Scientific Initiation
• Start date: December 07, 2024
• End date: March 06, 2025
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal Investigator: Ana Carolina Lorena
• Grantee: Lucas Ribeiro do Rêgo Barros
• Supervisor: Kate Smith-Miles
• Host Institution: Divisão de Ciência da Computação (IEC). Instituto Tecnológico de Aeronáutica (ITA). São José dos Campos , SP, Brazil
• Institution abroad: University of Melbourne, Australia
• Associated to the scholarship: 24/07637-9 - Analyzing meta-data from public repositories, BP.IC

Abstract

Machine learning (ML) approaches have become a vastly studied solution for challenging problems that are difficult to quantify or solve using traditional methods. One of the first challenges on solving a ML problem is determining which algorithm best fits its instance domain. Since no single ML algorithm can achieve optimal performance across every classification problems, choosing the right algorithm requires careful evaluation of the unique characteristics of each dataset features. This type of analysis, based on finding the unique features from a dataset, is key to the Meta-Learning (MtL) approach which investigates how dataset features relate to the performance of ML algorithms. By doing so, it provides more information for selecting the most appropriate algorithm for a given dataset, helping to identify conditions where specific ML techniques are most effective or may underperform. To optimize algorithm selection, the Instance Space Analysis (ISA) toolkit, developed at The University of Melbourne (UNIMELB), offers a powerful framework for evaluating the performance of many algorithms across a wide range of problem instances. By analyzing the meta-features of these instances, ISA maps algorithm performance onto a 2D space, providing a more intuitive understanding of algorithm behavior. This project proposal aims to expand the current implementation of ISA by integrating the PYTHIA module into the PyISpace package. PYTHIA will analyze the structural properties of instances, enabling automated algorithm prediction and recommendation. This integration is expected to enhance the accuracy of algorithm selection, streamlining the process for users and ultimately advancing the field of algorithm selection in machine learning. In partnership with The University of Melbourne, this BEPE proposal aims to advance the development of automated methods for algorithm recommendation, optimizing the selection process. (AU)