Abstract

Abstract
Hierarchical document classification is essential for structuring large-scale textual corpora in domains such as digital libraries and academic repositories. While recent advances in large language models (LLMs) have opened new possibilities for text classification, their applicability to hierarchical settings under real-world constraints remains underexplored. This study investigates both generative and discriminative transformer-based models, evaluating their effectiveness across multiple inference strategies: zero-shot baseline, local fine-tuning, and a global approach using category-specific models. Experiments on two real-world hierarchical datasets provide a comprehensive comparison of classification accuracy, F1-macro scores, and inference times. The results highlight that, although generative LLMs can deliver competitive (yet variable) performance at higher levels of the hierarchy, their high inference costs hinder their use in time-sensitive applications. In contrast, fine-tuned discriminative models—particularly BERT-based architectures—consistently offer a more favorable trade-off between performance and efficiency. © 2025 Elsevier B.V., All rights reserved.

Abstract
This research addresses the challenge of generating synthetic data that resembles real-world data while preserving privacy. With privacy laws protecting sensitive information such as healthcare data, accessing sufficient training data becomes difficult, resulting in an increased difficulty in training Machine Learning models and in overall worst models. Recently, there has been an increased interest in the usage of Generative Adversarial Networks (GAN) to generate synthetic data since they enable researchers to generate more data to train their models. GANs, however, may not be suitable for privacy-sensitive data since they have no concern for the privacy of the generated data. We propose modifying the known Conditional Tabular GAN (CTGAN) model by incorporating a privacy-aware loss function, thus resulting in the Private CTGAN (PCTGAN) method. Several experiments were carried out using 10 public domain classification datasets and comparing PCTGAN with CTGAN and the state-of-the-art privacy-preserving model, the Differential Privacy CTGAN (DP-CTGAN). The results demonstrated that PCTGAN enables users to fine-tune the privacy fidelity trade-off by leveraging parameters, as well as that if desired, a higher level of privacy.

Abstract
Our team previously introduced an innovative concept for an "invisible" Electrocardiography (ECG) system, incorporating electrodes and sensors into a toilet seat design to enable signal acquisition from the thighs. Building upon that work, we now present a novel dataset featuring real-world, single-lead ECG signals captured at the thighs, offering a valuable resource for advancing research on thigh-based ECG for cardiovascular disease assessment. To our knowledge, this is the first dataset of its kind. The tOLIet dataset comprises 149 ECG recordings collected from 86 individuals (50 females, 36 males) with an average age of 31.73 ± 13.11 years, a mean weight of 66.89 ± 10.70 kg, and an average height of 166.82 ± 6.07 cm. Participants were recruited through direct contact with the Principal Investigator at Centro Hospitalar Universitario de Lisboa Central (CHULC) and via clinical consultations conducted at the same institution. Each recording includes four differential signals acquired from electrode pairs embedded in the toilet seat, with reference signals obtained from a standard 12-lead hospital ECG system.

Abstract
Traditional machine learning techniques assume that data is drawn from a stationary source. This assumption is challenged in contexts with data streams for presenting constant and potentially infinite sequences whose distribution is prone to change over time. Based on these settings, detecting changes (a.k.a. concept drifts) is necessary to keep learning models up-to-date. Although state-of-the-art detection methods were designed to monitor the loss of predictive models, such monitoring falls short in many real-world scenarios where the true labels are not readily available. Therefore, there is increasing attention to unsupervised concept drift detection methods as approached in this paper. In this work, we present an unsupervised and interpretable method based on Radial Basis Function Networks (RBFN) and Markov Chains (MC), referred to as RMIDDM (Radial Markov Interpretable Drift Detection Method). In our method, RBF performs, in the intermediate layer, an activation process that implicitly produces groups of observations collected over time. Simultaneously, MC models the transitions between groups to support the detection of concept drifts, which happens when the active group changes and its probability exceeds a given threshold. A set of experiments with synthetic datasets and comparisons with state-of-the-art algorithms demonstrated that the proposed method can detect drifts at runtime in an efficient, interpretable, and independent way of labels, presenting competitive results and behavior. Additionally, to show its applicability in a real-world scenario, we analyzed new COVID-19 cases, deaths, and vaccinations to identify new waves as concept drifts and generate Markov models that allow understanding of their interaction.

Abstract
The importance of time series forecasting drives continuous research and the development of new approaches to tackle this problem. Typically, these methods are introduced through empirical studies that frequently claim superior accuracy for the proposed approaches. Nevertheless, concerns are rising about the reliability and generalizability of these results due to limitations in experimental setups. This paper addresses a critical limitation: the number and representativeness of the datasets used. We investigate the impact of dataset selection bias, particularly the practice of cherry-picking datasets, on the performance evaluation of forecasting methods. Through empirical analysis with a diverse set of benchmark datasets, our findings reveal that cherry-picking datasets can significantly distort the perceived performance of methods, often exaggerating their effectiveness. Furthermore, our results demonstrate that by selectively choosing just four datasets - what most studies report - 46% of methods could be deemed best in class, and 77% could rank within the top three. Additionally, recent deep learning-based approaches show high sensitivity to dataset selection, whereas classical methods exhibit greater robustness. Finally, our results indicate that, when empirically validating forecasting algorithms on a subset of the benchmarks, increasing the number of datasets tested from 3 to 6 reduces the risk of incorrectly identifying an algorithm as the best one by approximately 40%. Our study highlights the critical need for comprehensive evaluation frameworks that more accurately reflect real-world scenarios. Adopting such frameworks will ensure the development of robust and reliable forecasting methods.