Fabio Porto

Heteroscedastic time series exhibit non-constant variance over time, which undermines the effectiveness of classical anomaly detection methods based on fixed thresholds. In particular, it affects the identification of contextual anomalies whose detection depends on local variability. To address this challenge, this paper proposes the resilient transformation (RT), which modifies the original series through complementary ensemble empirical mode decomposition, selection of high-frequency components based on roughness, differentiation, and normalization by local dispersion. This transformation equalizes variance and highlights pointwise deviations. Evaluations using the Yahoo! S5 dataset, with the Harbinger framework, show that RT improves the performance of classical methods and can be combined with thresholding to form an autonomous detector ( detection method ) named anomaly detector by resilient transformation.

Debugging machine learning models is essential to improving their robustness and performance. This work explores a data-driven debugging approach based on problematic data regions—subsets of the data where the model performs poorly compared to others. These regions often reflect problems such as class imbalance, bias, or unfairness. We propose to improve model performance by focusing primarily on the data, using a specialized algorithm to identify problematic regions in the datasets, and then applying targeted interventions. A common cause for poor performance is class imbalance within a problematic region. For such scenarios, we apply data augmentation in a controlled manner, avoiding excessive introduction of synthetic data. Our experiments demonstrate that it is possible to improve model performance by focusing exclusively on problematic data regions rather than the entire dataset.

O aumento da complexidade em aplicações de aprendizado de máquina exige sistemas que garantam rastreabilidade e reprodutibilidade. Este trabalho apresenta a abordagem Twinscie-Prov, uma adequação do padrão W3C PROV para estruturar a proveniência de dados e processos ao longo do ciclo de vida de Machine Learning (ML) no sistema Twinscie. Os dados de proveniência são armazenados no sistema NoSQL Neo4j, permitindo consultas complexas e auditoria. Estudos preliminares mostram que, para consultas envolvendo navegação no grafo de dependências, típicas em dados de proveniência, a implementação no Neo4j é até cinco ordens de grandeza mais rápida que a baseada em logs.

Time series prediction is critical in domains such as economics, industry, and agriculture. Real-world scenarios often involve challenges like low data frequency, high variability, and non-repetitive patterns. Traditional statistical models and machine learning approaches, including Long Short-Term Memory (LSTM) networks, underperform in low-data contexts due to overfitting risks, intensive training requirements, and the lack of benchmarks tailored to such scenarios. Large language models (LLMs) have emerged as tools for time series forecasting, leveraging their ability to generalize and capture temporal dependencies and patterns across datasets without requiring extensive task-specific feature engineering. This study investigates the potential of time series foundation models (TSFMs), specifically Lag-Llama and Chronos, on low-frequency datasets by comparing zero-shot prediction across one-step-ahead and multi-step-ahead approaches. Our findings evaluate predictive error, robustness, efficiency, and applicability, showing how TSFMs address these limitations and enhance forecasting in data-scarce scenarios.

This brief communication presents the deployment of open-source large language models (LLMs) on the Santos Dumont Brazilian Supercomputer to support the Laborat´orio Nacional de Computa¸c˜ao Cient´ıfica (LNCC) academic community. The solution, named Carcar´a, leverages key optimizations such as dynamic quantization techniques to enable an accessible, scalable, and cost-effective implementation. As a result, a single-instance stateof- the-art independent model was deployed on a single node equipped with 4 × NVIDIA H100 GPUs. The quantized models fit entirely within the node’s VRAM, enabling horizontal scaling without the need for inter-node synchronization. In this way, we could provide access to these models to the entire LNCC academic community using only four computational nodes, demonstrating the efficiency and scalability of this approach. Crucially, our approach ensures data sovereignty, allowing LNCC researchers and postgraduate students to utilize AI for sensitive research topics with full control over their data, free from the privacy risks associated with proprietary solutions. This initiative strengthens national scientific autonomy while providing secure and efficient AI tools for academic and research advancements. The code to deploy this solution is openly available, and we encourage other institutions to adapt it and support their own communities. Discussions with the Brazilian Ministry of Science, Technology and Innovation are underway to expand this strategic solution to other research centers and universities in Brazil.