BASWE: Balanced Accuracy-based Sliding Window Ensemble for classification in imbalanced data streams with concept drift

• Ano de defesa: 2023
• Autor(a) principal: Oliveira, Douglas Amorim de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Concept Drift; Ensemble; Aprendizado de Máquina; Dados Desbalanceados; Imbalanced Datasets; Machine Learning
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/100/100131/tde-01012024-172026/

In the wake of the exponential growth in data generation witnessed in recent decades, this thesis addresses the inherent challenges of classification tasks within data streams, characterized by their continuous, real-time flow and inherent dynamic nature. With a focus on the compounding issues of concept drift and class imbalance, the thesis introduces the Balanced Accuracy-based Sliding Window Ensemble (BASWE), a novel ensemble method to deal with these challenges. BASWE leverages Balanced Accuracy, sliding windows, and resampling techniques to effectively handle imbalanced classes and concept drifts, ensuring robust performance even as data patterns evolve. In experiments conducted on 40 datasets, comprising 16 real-world datasets and 24 synthetic datasets generated under three configurations - no drift, gradual drift, and sudden drift - and with varying imbalance ratios, BASWE demonstrated superior performance compared to seven other state-of-the-art algorithms (CALMID, CSARF, KUE, OOB, ROSE, SMOTE-OB, and UOB) in terms of F1 Score and the Kappa statistic.