Studying meta-features at an instance-level: studies on unstructured datasets

• Grant number: 25/19506-9
• Support Opportunities: Scholarships abroad - Research Internship - Scientific Initiation
• Start date: December 01, 2025
• End date: February 28, 2026
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Ana Carolina Lorena
• Grantee: Thiago Galante Pereira
• Supervisor: Telmo de Menezes e Silva Filho
• Host Institution: Divisão de Ciência da Computação (IEC). Instituto Tecnológico de Aeronáutica (ITA). Ministério da Defesa (Brasil). São José dos Campos , SP, Brazil
• Institution abroad: University of Bristol, England
• Associated to the scholarship: 25/10304-4 - Studying meta-features at an instance-level, BP.IC

Abstract

Building efficient machine learning models for new datasets remains a challenging and resource-intensive process, as it requires exploring vast spaces of algorithms, architectures, and hyperparameters. Meta-learning (MtL) has emerged as a promising approach to mitigate this challenge by relating dataset characteristics to algorithm performance, thereby enabling more informed model selection. While traditional meta-learning studies operate at the dataset level, recent advances highlight the potential of instance-level analysis, where each data point is individually characterized to capture local complexity and predictive difficulty. This project focuses on the study of instance-level meta-features for unstructured datasets, particularly images, through embeddings generated by diverse deep learning architectures, such as Convolutional Neural Networks. By systematically extracting embeddings from multiple model layers and applying instance hardness measures with toolkits such as PyHard, the research investigates how architectural inductive biases shape the embedding space and affect meta-feature quality. The results are expected to provide deeper insights into the relationship between data representation and algorithm behavior, fostering the development of more accurate, cost-efficient, and automated model selection strategies. In addition to its scientific contributions, the project promotes collaboration between the Instituto Tecnológico de Aeronáutica (Brazil) and the University of Bristol (UK), strengthening research ties in meta-learning, data complexity analysis, and automated machine learning. The knowledge produced will be disseminated through open-science practices, enabling reproducibility and encouraging broader adoption of instance-level meta-features for analyzing unstructured data.