Category: Defenses

Dissertation (January 12, 2026): Edson Paulo da Silva Pinto Sobrinho

Student: Edson Paulo da Silva Pinto Sobrinho

Title: Fine-tuning detection criteria to enhance anomaly identification in time series

Advisors: Eduardo Soares Ogasawara (advisor) and Kele Teixeira Belloze (co-advisor)

Committee: Eduardo Soares Ogasawara (CEFET/RJ), Kele Teixeira Belloze (CEFET/RJ), Rafaelli de Carvalho Coutinho (CEFET/RJ) and Fábio André Machado Porto (LNCC)

Day/Hour: January 12, 2026 / 2 p.m.

Room: https://teams.microsoft.com/meet/24688731415374?p=YVHyDYwIW66rCysKq0

Abstract: Anomaly Detection (AD) is the problem of identifying observations that do not conform to typical ones in a time series. Detection methods implicitly define detection criteria, such as deviation measures, filter thresholds, and candidate anomaly selection strategies. Choosing inappropriate criteria results in inaccurate outputs, generating spurious alerts or missing events. Adjusting these criteria is essential for monitoring systems. To address this challenge, this study explores the fine-tuning of deviation measures, filter thresholds, and candidate selection strategies. Experimental results show that the proper choice of criteria significantly improves AD performance, often with greater impact than changing the detection methods.

Dissertation (January 08, 2026): Luiz Cláudio Lemos de Oliveira

Student: Luiz Cláudio Lemos de Oliveira

Title: Motif Detection in Time Series Using Autoencoders: An Analysis of Their Application to ECG Data

Advisor: Eduardo Soares Ogasawara

Committee: Eduardo Soares Ogasawara (CEFET/RJ), Laura Silva de Assis (CEFET/RJ), Helga Dolorico Balbi (CEFET/RJ) and Rebecca Pontes Salles (INRIA/FRA)

Day/Hour: January 08, 2026 / 10 a.m.

Room: https://teams.microsoft.com/meet/2899124353229?p=ykqgR5NeTJfPfXOhSF

Abstract: The discovery of motifs in biomedical time series, such as electrocardiograms (ECGs), involves identifying recurrent patterns that may contain valuable diagnostic information. Traditional methods, such as SAX, are limited by strong statistical assumptions, which are particularly inadequate for complex physiological signals. In parallel, autoencoders have demonstrated superior ability to learn nonlinear representations, but their application to motif discovery in ECG data remains unexplored, constituting a significant methodological gap. This work proposes a framework that replaces SAX discretization with neural encoding while preserving the discovery pipeline based on Shannon’s entropy and frequency of occurrence. The methodology was developed in three stages: (i) validation of the autoencoder’s reconstruction ability, (ii) training models with data from the MIT-BIH Arrhythmia Database, and (iii) systematic experimental comparison with the traditional SAX method through detection experiments, parametric sensitivity analysis, and evaluation of generalization capacity. It is concluded that replacing traditional discretization with neural encoding is feasible and provides quantitative and qualitative gains in motif discovery in ECG signals, establishing a methodological basis for developing automated biomedical signal analysis tools.

Dissertation (December 18, 2025): Michel Siqueira Reis

Student: Michel Siqueira Reis

Title: Matching Detections to Events in Time Series with Computational Efficiency and Guaranteed Optimality

Advisors: Rafaelli Coutinho (Advisor) and Eduardo Ogasawara (Co-advisor)

Committee: Rafaelli Coutinho (Cefet/RJ), Eduardo Ogasawara (Cefet/RJ), Laura Assis (Cefet/RJ) and Rebecca Salles (INRIA)

Day/Hour: December 18, 2025 / 1 p.m.

Room: https://teams.microsoft.com/l/meetup-join/19%3ae20c8697654543fc9dd1e9924de5c2c0%40thread.tacv2/1763159776096?context=%7b%22Tid%22%3a%228eeca404-a47d-4555-a2d4-0f3619041c9c%22%2c%22Oid%22%3a%2254af42a0-5f30-4905-ac8d-10b96c6db26b%22%7d

Abstract: This work presents SmartSoftED, an optimized metric for evaluating the detection of point events in time series. The original metric, SoftED, introduces a “soft” evaluation based on temporal tolerance, assigning gradual scores to detections that occur near actual events. However, its current formulation relies on a greedy approach that does not guarantee optimality in all cases and incurs a quadratic computational cost, limiting its applicability in large-scale or real-time processing environments. SmartSoftED overcomes these limitations by introducing a strategy that decomposes the problem into manageable, disjoint subproblems: some can be solved efficiently without loss of optimality, while others are modeled as maximum-weighted matching problems on unbalanced bipartite graphs. This approach preserves the optimality of correspondences between detections and events while significantly reducing computational cost. In practice, the method achieves an average speedup of two orders of magnitude, making it suitable for certain large-scale applications and systems with strict temporal constraints.

Dissertation (December 8, 2025): Fernando Henrique de Jesus Fraga da Silva

Student: Fernando Henrique de Jesus Fraga da Silva

Title: Aprendizado por Reforço Profundo Aplicado à Negociação Intradiária de Múltiplas Ações

Advisors: Eduardo Bezerra da Silva (advisor) and Pedro Henrique González Silva (co-advisor)

Committee: Eduardo Bezerra da Silva (Cefet/RJ), Pedro Henrique González Silva (UFRJ), Aline Marins Paes Carvalho (UFF) e Glauco Fiorott Amorim (Cefet/RJ)

Day/Hour: December 8, 2025 / 3 p.m.

Room: https://teams.microsoft.com/v2/?meetingjoin=true#/l/meetup-join/19:PKOJTuK7mfHSDE6QkCWQCYp71f0xOMNoRgSUj4wjMKc1@thread.tacv2/1763760050816?context=%7b%22Tid%22%3a%228eeca404-a47d-4555-a2d4-0f3619041c9c%22%2c%22Oid%22%3a%22c03d6068-4733-48a6-bbb4-aa78f351d9cf%22%7d&anon=true&deeplinkId=91733be2-9804-4f09-ac6a-f1a362e67de8

Abstract: The stock market is a dynamic and volatile environment in which publicly traded companies negotiate fractions of their value, subject to continuous price fluctuations influenced by economic, political, and social factors. Anticipating these fluctuations is a complex task, especially in the context of intraday trading, where buy and sell decisions must be made within very short time intervals based on rapidly changing data. In this scenario, Reinforcement Learning (RL) emerges as a promising paradigm capable of developing adaptive strategies through the continuous interaction between agent and environment. This dissertation investigates the use of Deep Reinforcement Learning (DRL) techniques in financial trading, focusing on intraday scenarios involving multiple stocks. It proposes a DRL-based approach to estimate buy and sell actions simultaneously across various assets, using high-granularity market data to better approximate real trading conditions. Experimental analyses were conducted using the Proximal Policy Optimization (PPO) algorithm. The results indicate that the proposed agent outperformed traditional benchmark strategies, achieving gains exceeding 10 percentage points in certain cases.