This work presents the development of Lucas, a virtual assistant built entirely with open-source tools to support older adults in their daily routines. Built upon the Robot Operating System (ROS 2), the system is composed of modular nodes for wake word detection, speech processing, and natural language interaction. The assistant is triggered by the spoken expression “Opa Lucas” after which it records the user’s voice, transcribes the audio to text using Whisper, and processes the instruction using the LLaMA 3 Large Language Model. Depending on the user’s request, the system may access external tools, such as reminders and weather forecasts, via predefined APIs and SQL queries. The entire interaction is presented through a multimodal interface built with Pygame, combining speech synthesis and animated feedback. The assistant supports speaker identification through voice verification. To evaluate the assistant, a set of instructions was proposed to assess the accuracy, errors made, and the nature of the responses, including aspects such as harmfulness, helpfulness, and coherence. Additionally, the assistant was reviewed and tested by a geriatric care professional. The results demonstrate the assistant’s potential to deliver coherent, accurate, and timely instructions, suggesting its applicability in daily routines. The assistant’s source code is publicly available in a GitHub repository, along with documentation for setup and usage.
Accurate identification of coma etiology is essential for effective treatment in neurocritical care. However, significant challenges persist, especially in settings with limited resources. This study investigates the potential of electroencephalogram (EEG) signals from distinct brain regions for predicting coma etiology using an adapted Convolutional Neural Network combined with Long Short-Term Memory (CNN-LSTM) model. Building on prior hybrid CNN-LSTM architectures that processed all EEG channels collectively, we refined the model to integrate raw EEG data from specific regions (frontal, central, parietal, temporal, and occipital) with patient-specific clinical features (age, gender) and statistical EEG metrics. The model was evaluated on a real-world dataset of 50 patients with coma etiologies including traumatic brain injury, metabolic coma, and stroke, collected at the Clinical Hospital of the Federal University of Uberlândia (HC-UFU). Experimental results demonstrate that the frontal region, when combined with clinical features, yielded the highest predictive performance, achieving an accuracy of 70
Timely disease diagnosis is essential for reducing livestock mortality, as delays caused by reliance on human visual inspection can lead to severe consequences. Early detection of illness hinges on access to detailed, individual animal data capable of capturing subtle behavioral and physiological changes before visible symptoms emerge. A robust classification system is therefore critical for automating behavior analysis, enabling prompt identification of deviations from normal patterns that may indicate illness and facilitating timely intervention. This work reports the development of an innovative livestock behavior classification and health monitoring device. The system incorporates a wireless neck collar equipped with motion and temperature sensors for continuous animal activity monitoring. A prototype collar was tested on two heifers in a controlled field setting under video surveillance. Collected data were manually annotated by human observers, classifying the heifers’ behaviors into five key categories: eating, drinking, walking, lying down, and standing. The study covers the entire process of creating the monitoring system, from the design and construction of the neck collar to data collection and annotation. It also details the complete data mining workflow, including data cleaning, pre-processing, and comparative analysis using six shallow Machine Learning (ML) algorithms. The results indicate that Gradient Tree Boosting is the best-performing model based on classification metrics appropriate for multiclass imbalanced data, such as the F1-score, Weighted F1-score, and Kappa coefficient. The designed device, integrated with a ML model, proved to be an effective and low-cost approach for classifying livestock behavior.
A significant portion of the textual corpus lacks the annotated labels needed for training supervised models. Manual labeling of such data is costly and impractical. In response to this challenge, semi-supervised learning (SSL) has emerged as a pivotal approach to harnessing the potential of both labeled and unlabeled data. This paper explores the integration of adversarial perturbation with SSL, particularly focusing on text classification tasks. We propose an innovative approach called APVAT, which incorporates adversarial perturbation into the Attention Virtual Adversarial Training (VAT) model. We performed experiments in five benchmark datasets, exploring the impact of different embeddings such as fastText, GloVe BERT, and GPT-2. Our contributions are twofold: First, our approach improves classification accuracy in datasets with little training data compared to previous methods, even when labeled data is scarce (e.g., 10
Dermatological diseases represent one of the leading causes of clinical consultations in dogs, affecting over 30
Solving machine learning problems involving interval-valued data is a challenging task that arises in various real-world applications, such as heart rate prediction and astronomical data analysis. Motivated by this, several nonlinear regression methods and classifiers have been proposed in recent years to address this specific data type. In this paper, we introduce two novel variants of the Minimal Learning Machine (MLM) adapted to interval-valued inputs, targeting both regression and classification tasks. The proposed variants explicitly model a direct dependency between input and output intervals: the lower (upper) bound of the output is predicted based on both bounds of the input. For regression tasks, we compare our models against seven state-of-the-art nonlinear approaches: three Extreme Learning Machine (ELM) variants for interval-valued data, two kernel regression extensions, without explicitly modeling interval dependencies. For classification, we evaluate three interval-based logistic regression models. Experimental results on synthetic datasets with varying configurations and on real-world data demonstrate that our MLM variants achieve comparable or superior performance. These findings highlight the effectiveness of the proposed methods and their potential as competitive alternatives for learning from interval-valued data.
This paper presents a study of artifacts left in images by the generative process of facial deepfakes. We hypothesize that, even when facial regions are blurred or erased to protect privacy and biometric data, an image can still be classified as fake. To address such privacy-critical scenarios, our framework is specifically designed to reliably detect facial deepfakes, even when all facial regions are deliberately obfuscated to preserve individual privacy. Our approach is grounded in the observation that spatial artifacts, high-frequency components, and region-localized features in both the pixel and transformed domains serve as strong evidence of deepfake forgeries. To capture these cues, we extract three representations of each image - the raw pixel domain, wavelet based Multi-Resolution Analysis (MRA), and the Short-Time Fourier Transform (STFT) - and process them using a state-of-the-art pre-trained feature extractor, the as OpenAI’s CLIP embedding model. The generated embeddings are used to train two different classification methods (Logistic Regression and Multi-Layer Perceptron based) and we verify if the extra information presented to the feature extractor (STFT and MRA) improves deepfake identification metrics. Our contribution can be seen as a systematic approach to enhance deepfake classifier accuracies in information deprived scenarios, where the identity of a given subject is hidden by erasing or distorting its facial landmarks. We evaluate our framework on the DFFD deepfake dataset with some pre-processing of the database (to compute the transforms and hide facial identity) and we demonstrate promising detection performance under these conditions.
Large language models (LLMs) have gained a lot of visibility recently and have been widely used. They revolutionized the field of artificial intelligence by allowing machines to process and reproduce human-like texts in an unprecedented way. In the same way that this technology promotes the creation of innovative applications, from chatbots and virtual assistants to content creation tools and personalized recommendation systems, it also faces major challenges and risks, such as adversarial attacks. Adversarial attacks aim to reveal the vulnerability of deep neural network models, the basis of LLMs, enabling the development of defense methods and thus making these models more robust. This research sought to investigate the robustness of LLMs fine-tuned for the Portuguese language against adversarial attacks, specifically the BERTimbau and Sabiá models. Experimental results demonstrate an attack success rate against the BERTimbau model of approximately 93
The increasing importance of digital health and the need for better health literacy require effective methods to access and understand Electronic Health Records (EHRs). While Large Language Models (LLMs) show promise in this domain, traditional Retrieval-Augmented Generation (RAG) struggles to handle the complex, interconnected nature of clinical data. GraphRAG emerges as a powerful alternative, leveraging knowledge graphs (KGs) to capture semantic relationships within EHRs. This research evaluates the effectiveness of graph expansion in a GraphRAG to enhance information retrieval from FHIR-formatted medical data. We propose a 1-hop expansion approach built upon a lexical search baseline which, while inheriting some limitations of traditional keyword-based retrieval, significantly enhances LLMs’ access to comprehensive and diverse contextual information. Our evaluation, using synthetic patient data and a targeted set of questions across five models, reveals that the 1-hop expansion strategy consistently outperforms the baseline in subjective metrics like comprehensiveness and diversity, and frequently in quantitative metrics such as answer and contextual relevancy. These results highlight the potential of our proposal to enhance LLM response when querying intricate medical data.
Clustering plays a fundamental role in attributed graphs, which incorporate both topological structure and node attributes represented as feature vectors. Deep clustering methods based on Graph Neural Networks (GNNs) have proven effective in extracting patterns from such data. Most existing approaches use a traditional clustering algorithm to identify representative elements, which are later employed to train a GNN and, finally, the clustering task. However, when selecting representative elements, these clustering algorithms consider only the feature vector of each instance, neglecting topological information. This limitation negatively impacts the GNN learning process. To address this issue, we propose Deep Graph Clustering via Graph Neural Network and Seed Detection (DGCSD), a model consisting of three modules: (1) the seed detection module, which identifies representative nodes; (2) the embedding module, which employs a graph attentional network to capture global topological information; and (3) the self-supervised module, which leverages the representative nodes to guide the clustering task. An advantage of our algorithm is that it integrates both information, the topological structure and node attributes across all modules to identify representative elements. This is the first GNN-based clustering algorithm that incorporates seed detection, establishing a significant reference for future research. The empirical analysis of real-world graphs provides evidence that combining a seed detection algorithm with a GNN model is competitive compared to well-established algorithms.
Modeling complex systems from irregular multivariate time series (MTS), common in engineering and scientific domains, presents significant challenges due to asynchronous measurements, varying sampling rates, and missing data. Building upon architectures that explicitly handle temporal irregularities using time-informed recurrent units, this work introduces a refined approach that fundamentally decomposes the latent state update process. We propose separating the function responsible for incorporating new measurements from the function governing the temporal evolution of the hidden state. Specifically, the temporal translation over a given time delta is modeled using a dedicated transformation parameterized by a matrix constrained to be near-orthogonal. A separate ingestion function updates this time-evolved state with new measurement information when available. This explicit decomposition allows for a more principled representation of continuous-time dynamics, acting as a general regression framework applicable to both imputation and forecasting tasks. We demonstrate through experiments that our model is competitive with state-of-the-art techniques, while providing the desirable properties of near-orthogonal time translations.
Time series snippet discovery aims to summarize complex sequences by extracting representative subsequences that capture predominant behaviors. While Matrix Profile-based methods like Snippet-Finder offer strong robustness, they face significant computational challenges with long-duration or high-resolution time series, often requiring O(M^2) operations. We propose RS4 (Restricted Search Space for Snippet Selection), a hybrid method that integrates clustering techniques with a refined Matrix Profile Distance search strategy. Our approach first segments and normalizes subsequences, applies clustering to identify cohesive groups, selects medoids as candidate snippets, and computes a restricted MPdist profile for final selection. This approach reduces the search space complexity to O(Mn) , where n is the number of clusters, without sacrificing pattern fidelity. Experimental evaluation on the MixedBag dataset and long-duration sleep recordings demonstrates that RS4 achieves a 78
Traditional assessments of metaheuristics typically focus on performance metrics when solving benchmark problems, with limited attention to the underlying mechanisms that drive collective intelligence. This performance-centric approach fails to indicate why certain metaheuristics are successful in specific scenarios and provides minimal guidance for improving their design. To better understand these inner dynamics, we conducted a multidimensional analysis of the swarm dynamics in a Reinforcement Learning-based metaheuristic that can combine search behaviors from different Swarm Intelligence metaheuristics. Using Interaction Networks, Search Trajectory Networks, Improvement Frequency, and Population Turnover metrics simultaneously, we reveal how the RL approach dynamically switches between different swarm behaviors depending on the problem. Our results demonstrate that the RL-based metaheuristic matches the performance of the most effective specialised algorithm for each benchmark function by adaptively mimicking its social interaction patterns, search trajectories, and fitness development. We show that for the F1 function, the RL approach adopts behaviors similar to GWO, while for the F2 function, it transitions to GPSO-like patterns, aligning with the expected best-performing characteristics reported in the literature. This multidimensional approach provides insights into how adaptive metaheuristics combine different behaviours at optimisation stages, suggesting a pathway toward self-adaptive metaheuristics that automatically select optimal search strategies according to the simulation scenario.
The development of domain-specific language models faces significant challenges due to the scarcity of labeled data, especially in low-resource languages such as Portuguese. Annotating data is expensive and time-consuming, limiting the ability to train effective models in specialized contexts. To address this, we investigate a self-supervised fine-tuning strategy based on the BERTimbau pre-training protocol. This approach allows the model to improve generalization using only unlabeled data, avoiding the need for manual annotation. We explore different combinations of unfrozen layers and learning rate configurations to identify training regimes that balance performance and computational cost. The method is evaluated on three sentiment analysis datasets in Portuguese, each from a distinct domain. Results show that unfreezing only the final layer, together with a properly tuned learning rate, achieves performance comparable to traditional fine-tuning approach. These findings confirm the method’s viability in low-resource settings and its potential to scale to large unlabeled datasets. The approach provides an efficient alternative for adapting language models when annotated data is limited.
We present a theoretical result showing that a learning algorithm can be seen to be an approximation of Solomonoff induction. A constant multiplicative bound between the predictive probability of a learning algorithm and the Solomonoff semi-distribution characterises this approximation. Our main contribution is to state and prove this result concisely, in general, clarifying the theoretical role of algorithmic probability in practical learning. We also establish a positive correlation between the Vapnik Chervonenkis (VC) dimension of a model class and the Kolmogorov complexity of individual hypotheses within it, reinforcing the conceptual alignment between learnability and compressibility. We contextualise these findings by discussing implications on the MDL principle and the emergent behaviour of large language models (LLMs), which empirically approximate algorithmic prediction. An example and comparisons with PAC learning, Bayesian inference, and reinforcement learning are provided.
This scoping review explores the use of robotic technologies as intervention tools for children with Autism Spectrum Disorder (ASD), an area of growing interdisciplinary interest. Following the PRISMA-ScR guidelines, we conducted a comprehensive search in the Scopus, IEEE Xplore, and ACM Digital Library databases for studies published between 2018 and 2025. From an initial set of 1167 publications, 27 articles met the inclusion criteria and were analyzed in detail. The results demonstrate a diverse range of robotic applications, with the NAO robot featured in 13 out of the 27 selected studies. Robotics interventions were primarily implemented in assistive therapy, social skills training, educational settings, and joint attention initiatives, predominantly targeting children aged 5 to 10 years. The volume of relevant publications peaked in 2020 and declined in subsequent years, highlighting both increased early momentum and the current need for renewed research efforts in this domain. Overall, the review identifies critical research gaps, particularly in terms of personalized and interdisciplinary intervention approaches. While robotic technologies show considerable promise in supporting autistic children, further high-quality studies are required to assess their long-term effectiveness and to inform the development of individualized, evidence-based practices.
The growing challenges posed by infectious diseases, antimicrobial resistance, increasing cancer prevalence, and untreatable rare genetic conditions highlight the necessity to accelerate drug discovery. This process remains expensive and time-consuming but computational approaches like molecular docking combined with machine learning (ML) can impact on this process reducing costs and time. In molecular docking, the interaction between drug candidates (ligands) and target receptors is evaluated with Scoring Fuctions (SFs). The accurate prediction of the binding free energy in protein-ligand complexes performed by SFs remains a critical challenge in computational drug discovery. This study proposes CRRF-Score (Cumulative Ranking Random Forest Scoring Function), a ML-Based SF that integrates feature selection and ensemble models to enhance binding affinity prediction. CRRF-Score combines descriptors from DeltaVinaRF20, AutoDock Vina, and RDKit 2D, employing a diverse set of algorithms - Random Forest, Decision Tree, Lasso, Principal Component Analysis and Automatic Relevance Determination Regression for robust feature ranking. Our approach is trained on PDBbind 2018 and evaluated using CASF-2016 benchmarks, demonstrating competitive performance in Scoring Power and Ranking Power. Results show that our cumulative ranking strategy for feature selection combined with random forest algorithm for training the SF matches state-of-the-art commercial SFs in key metrics, offering a promising solution for molecular docking.
The electrocardiogram (ECG) is a non-stationary signal used to assess heart health and to detect systemic conditions such as mental stress and drug toxicity. However, labeling ECG data is notably costly and can limit the applicability of supervised deep neural networks in tasks involving this physiological signal. Self-supervised learning (SSL) is a two-stage approach that has gained attention in this scenario due to its labeling efficiency and knowledge transfer capabilities in various data modalities. However, recent studies describe the difficulties in transferring knowledge across different time series datasets. This work investigates several data-oriented strategies for pretraining self-supervised learning models in a cross-dataset setting to address a challenging ECG classification task. We introduce a data-driven regularization approach that perturbs the pretrained model using signals from multiple domains. We evaluated three well-established SSL models as backbones with different classification architectures across over 100 experiments, achieving superior performance to traditional SSL methods and a state-of-the-art supervised deep learning classification model.
Time series forecasting has significantly benefited from ensemble models, which consistently outperform individual approaches by combining predictions from diverse methodologies. This paper proposes a novel ensemble framework based on the copula formalism, enabling dynamic selection of copula functions tailored to the characteristics of each dataset. Our approach integrates three widely recognized forecasting models: ARIMA, Multilayer Perceptron (MLP), and Long Short-Term Memory (LSTM) neural networks and applies them to five key financial time series: Petrobras, Google, Nasdaq, S P500, and the GBT to USD exchange rate. The ensemble combines predictions using multiple copula functions and compares their performance against traditional ensemble methods, such as simple average, simple median, and MLP-based combination. Experimental results show that the proposed system achieves mean absolute percentage errors (MAPE) below 0.71, outperforming standard ensemble strategies. This improvement highlights the effectiveness of copula-based modeling as an adaptive and flexible tool for predictive fusion. We conclude that incorporating copula theory into ensemble learning brings substantial advancements in time series forecasting, as it allows dynamic model aggregation according to data behavior. The proposed method bridges statistical foundations and artificial intelligence, offering promising results in financial applications and potential for broader use across various forecasting domains.
Depression screening through social media has emerged as a promising research avenue; however, the scarcity of annotated data remains a significant barrier to effective model training. In this work, we evaluate several textual data augmentation strategies for screening users with depression through Brazilian-Portuguese Instagram posts. We explore three techniques of increasing complexity: simple post duplication, contextual word substitution using BERT-based models, and synthetic posts generation via Large Language Models (LLMs), both with and without the modulation of psychometric data from the Beck’s Depression Inventory (BDI-II) to create new instances. Experiments were conducted under both Single-Instance Learning (SIL) and Multiple-Instance Learning (MIL) frameworks, using multilingual sentence embeddings and an XGBoost classifier. Results reveal statistically significant differences among augmentation strategies, with LLM-based generation without BDI-II modulation achieving the highest performance. Contextual substitution proved to be a competitive and computationally efficient alternative. In contrast, psychometric modulation reduced model effectiveness, suggesting that artificially aligning emotional tone may compromise data quality. These findings underscore the importance of semantically coherent augmentation for sensitive applications in mental health. Code and supplementary material: https://github.com/Matheusutino/depression-data-augmentation .
