Dynamic Self-Organizing Maps for Multi-Label Classification in Continuous Data Streams

• Grant number: 26/03081-1
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: May 01, 2026
• End date: April 30, 2027
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal Investigator: Ricardo Cerri
• Grantee: Arthur Henrique Silva de Araujo
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Associated research grant: 22/02981-8 - Novelty detection in multi-label data streams classification, AP.PNGP.PI

Abstract

Classifying continuous data streams presents a significant challenge for machine learning algorithms. This is due to their non-stationary nature, the change and emergence of new concepts over time, and the potential for infinite label latency, meaning labels may never be available to update the classifier. The challenge of classifying these streams becomes even greater when examples can be assigned to multiple classes simultaneously, which characterizes a multi-label problem. In these problems, the change in the cardinality of example labels over time must also be considered. Although the literature contains works that address continuous multi-label data streams, few address the infinite-label latency problem. Those that do generally use the k-means clustering algorithm to build decision models, which requires the user to define an initial number of groups. In this research, we propose using dynamic self-organizing maps to eliminate the need for initial hyperparameters in classifier construction. The proposal aims to extend the SOM-AA method (An Algorithm Adaptation Method for Multi-Label Stream Classification using Self-Organizing Maps) published by our research group. The extension will allow the artificial neuron map to grow incrementally from the start of training, with new neurons inserted to represent new concepts or extensions of already known concepts. The method will be evaluated on synthetic and real datasets, compared with the literature, and assessed using specific evaluation measures. (AU)