Neural networks for out-of-distribution time series detection in one-class learning

• Ano de defesa: 2025
• Autor(a) principal: Júnior, José Gilberto Barbosa de Medeiros
• 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: Aprendizado de única classe; Aprendizado profundo; Deep learning; Neural networks; One-class learning; Redes neurais; Séries temporais; Time series
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-29072025-141957/

Time series classication presents unique challenges, particularly in domains where data labeling is expensive or where the class of interest is signicantly more prevalent than others. To address these challenges, one-class learning (OCL) emerges as an alternative, focusing exclusively on learning time series that belong to a single class, also referred to as the interest class. Existing OCL methods often suer from instability, reliance on counterexamples, and inadequate adaptations for capturing temporal dependencies. This dissertation proposes a novel OCL framework for time series, integrating neural network mechanisms and outof- distribution (OOD) detection techniques. To enhance temporal feature extraction, we introduce two new neural network components: (i) LeakySineLU, a novel activation function designed for time series tasks, and (ii) Deformable Convolutions for time series, which enable the capture of non-continuous and long-range dependencies between observations. These mechanisms are incorporated into TGNet, a proposed OCL method that utilizes Gaussian Mixture Models (GMMs) to model the distribution of the class of interest and identify out-of-distribution instances. Extensive experiments conducted on 112 datasets demonstrate that TGNet outperforms traditional OCL approaches, achieving a higher average F1-score in classication and, consequently, a higher mean ranking. Ablation studies conrm the individual contributions of the proposed mechanisms, reinforcing their role in advancing one-class learning for time series.