Extracting knowledge from digital social platforms is a new trend in software requirements engineering. The feedback generated by users in these environments is a very valuable source of needs, demands, dissatisfactions, and judgments about the software applications they use. The effectively processing of this volume of information for predicting the software product evolution is a great challenge, which can start from determining the most relevant contents, to the automatic elicitation of new requirements for that product. This paper proposes a method for generating software requirements from user feedback. This solution combines Deep Learning models for determining the relevant information, with an LLM (Large Language Model) to generate functional and non-functional requirements about a software and specific aspects from the feedback. This solution was evaluated using datasets of opinions from 4 different applications. The results were very promising, because they showed improvements over other reported solutions, and demonstrated that with a proper conception of information filtering and specification of interests through the prompt it is possible to generate good quality requirements using LLMs from an informal information source.
This paper presents a novel implementation of a Neural Network model in a low-power embedded system for burst detection in pipelines. It combines a quantized model with the LEA (Low-Energy Accelerator) and FRAM memory to perform signal processing and inference operations, optimizing the balance between accuracy and energy efficiency. The model was trained on a dataset of 187 signals (80 noise, 107 ruptures). Accuracy, precision, and recall metrics were used for model evaluation. After implementation in the microcontroller, a validation protocol was executed to assess the impact of quantization on the method’s accuracy.
Limited availability and class imbalance in tabular data make it difficult the development of effective models using machine learning methods. Synthetic data generation is a promising solution, but requires rigorous methodologies and extensive evaluation. This work proposes a methodology for generating and evaluating synthetic tabular data, especially from imbalanced datasets, illustrated with a case study in Intensive Care Units. The methodology comprises: (1) class balancing by oversampling, selecting the best technique with a quality metric; (2) synthetic data generation from the optimal balanced dataset, using SMOTE RSB* Adapted with Gaussian Noise, CTGAN and TVAE; and (3) a novel quality metric, TabDSFidelity, that integrates distributional similarity, correlation preservation, and predictive utility to guide the selection at each stage. Applied to ten datasets, the methodology demonstrated that the use of synthetic data selected by a quality metric significantly improves the performance of classification models compared to using only the original data. SMOTE RSB* Adapted with Gaussian Noise consistently generated the highest quality data according to TabDSFidelity in this study. It is concluded that the proposed methodology offers an effective framework for mitigating data sparsity and imbalance, facilitating the creation of more accurate and robust models.
In this paper a new embedded software package for efficiently computing the Continuous Wavelet Transform (CWT) in ARM Cortex cores, based on the original fast Continuous Transform (fCWT) algorithmic implementation is introduced. To execute the operations in ARM Cortex cores efficiently, the generic C++ implementation of the fCWT was adapted to a CMSIS-DSP based implementation in order to exploit the capabilities of the floating point unit (FPU) and the DSP instructions set with SIMD capabilities of ARM Cortex-M and Cortex-A cores. The package can be used for Continuous Wavelet Transform based feature extraction in edge computing signal processing applications that involve ARM based microcontroller families. A functional verification and a specific application oriented performance measurement approach in terms of real-time analysis ratio (RAR) for embedded processors is provided.
Accurate forecasting of the Consumer Price Index is crucial for economic stability, enabling policymakers to design effective monetary strategies and mitigate inflationary risks. Traditional models like ARIMA struggle to capture complex non-linear patterns and long-term dependencies in Consumer Price Index data, often leading to suboptimal predictive performance. To address these limitations, this study proposes a hybrid deep learning model combining a Deep Neural Network and a Transformer architecture. The Deep Neural Network extracts non-linear features, while the Transformer’s multi-head attention mechanism captures temporal dependencies, enhancing forecasting precision. The model incorporates advanced preprocessing techniques, including dynamic window normalization and outlier detection, ensuring robust data representation. Experimental results demonstrate superior performance, with a Mean Absolute Error (MAE) of 0.45, Root Mean Square Error (RMSE) of 0.62, and Mean Absolute Percentage Error (MAPE) of 1.8
Tattoo recognition and retrieval remain challenging tasks due to the diverse and intricate nature of tattoo designs. Existing approaches typically rely on visual features extracted from convolutional neural networks (CNNs), which may fail to capture the rich semantic information embedded in tattoos. In this paper, we propose a novel framework that integrates state-of-the-art visual features with textual descriptions generated by a multimodal large language model (MLLM). We explore the impact of different prompts on the quality of the generated captions and use CLIP to create textual embeddings. By combining cosine similarity scores from both modalities, our approach achieves superior performance in tattoo retrieval tasks. Experimental results demonstrate that our method outperforms traditional visual-only approaches, highlighting the importance of leveraging multimodal data for tattoo recognition. To the best of our knowledge, this is the first work to combine MLLM-generated textual descriptions with visual features for tattoo retrieval.
Scene text generation in images remains a major challenge for current text-to-image (T2I) systems, especially in morphologically rich and underrepresented languages such as Spanish. Existing generative models often produce text that is visually distorted, grammatically incorrect, or semantically inconsistent. In this work, we present a targeted approach to improve scene text generation in Spanish by fine-tuning a state-of-the-art diffusion model, FLUX.1-dev, using a novel dataset of Spanish-language memes (CCMD). Our methodology integrates prompt engineering, Low-Rank Adaptation (LoRA), and a custom evaluation protocol that includes human judgment. We demonstrate that increasing the number of denoising steps k leads to consistent improvements in legibility, alignment, and linguistic fidelity. Our results show that the proposed model outperforms baseline systems such as GPT-4 (DALL · E). We further propose a roadmap for building automatic evaluation frameworks that assess scene text not only lexically, but also semantically and visually, paving the way toward more inclusive and robust generative systems.
In Machine Learning, class imbalance is one of the most common challenges in classification tasks. This issue becomes even more pronounced in Natural Language Processing domains such as TASS, a Spanish-language emotion detection corpus characterized by a marked disproportion among categories. Traditional oversampling methods like SMOTE, based on k-nearest neighbors, lose effectiveness when applied to the high-dimensional spaces generated by modern language models. This work presents a probabilistic balancing framework that models the distribution of RoBERTa CLS embeddings (768 dimensions) using the covariance matrix estimated through the Ledoit–Wolf method, Lasso regression, and Elastic Net. From these distributions, realistic synthetic instances are generated for minority classes, drastically reducing the imbalance ratio without introducing semantic noise. The balanced embeddings are then classified using a lightweight multilayer perceptron (MLP), which eliminates the need for costly transformer fine-tuning. When evaluated on the TASS 2020 dataset, the best proposed algorithm achieved a Macro F1 score of 82.45
License Plate Recognition (LPR) in video sequences faces challenges such as motion blur, occlusions, and poor lighting. While most research focuses on detection and recognition, frame combination strategies remain mainly underexplored This paper introduces a novel segmentation-free plate recognizer that allows to estimate the confidence of each returned character together with an estimation of the network uncertainty. Evaluated on a challenging dataset of real-world Cuban plates with deformations and adverse conditions, we found combinations that outperforms traditional methods, achieving significant gains in accuracy for low-quality scenarios. These findings highlight the importance of temporal integration in video-based LPR systems.
This study addressed the problem of ambiguity in Software Requirements Specifications (SRS) through an innovative zero-shot classification approach, aiming to automatically identify and classify types of linguistic ambiguity without requiring large volumes of labeled data. We proposed a model based on Transformer and BERT architectures, adapted to capture ambiguous linguistic phenomena, utilizing two main components: an ambiguous expressions encoder and an ambiguity types encoder. The methodology included unsupervised pretraining with pseudo-labels and contrastive fine-tuning to maximize similarity between ambiguous expressions and their corresponding categories. Results demonstrated that the model with contrastive pretraining (CP) achieved an average performance of 41.78
Social security programs aim to protect vulnerable populations; however, accurately identifying individuals with significantly lower incomes than their peers (accounting for age, occupation, and education level) remains an operational challenge. This article proposes an innovative method for detecting economic vulnerability by combining income data enrichment with large language models in a multi-agent architecture, unsupervised clustering techniques, and statistical heuristics. The developed algorithm analyzes demographic and labor-related variables to estimate expected annual income by profile, thereby identifying atypical discrepancies that suggest vulnerability. This approach not only optimizes the prioritization of beneficiaries for targeted assistance but also serves as a preventive mechanism against the inadvertent exclusion of eligible groups. Preliminary results demonstrate the method’s effectiveness in detecting hidden vulnerability particularly among young adults aged 17–23, whose high underemployment rates ( ≈ 40% ) in recent national statistics closely align with the concentration of vulnerability detected. These findings underscore its potential as a complementary tool to enhance equity and efficiency in social policy implementation.
The integration of large language models (LLMs) with external sources has established Retrieval-Augmented Generation (RAG) as a key technology for improving reliability and reducing hallucinations. While RAG has proven effective, the paradigm has evolved to incorporate various techniques at different stages, aiming to further enhance its potential. However, a key challenge lies in identifying configurations of advanced techniques that improve performance. This study evaluates several technique configurations in a modular RAG system, with the objective of identifying those that deliver the best performance. Technologies such as LangChain, ChromaDB, and LLaMA 3.1:8B are employed. The evaluation focuses on metrics such as precision, recall, and F1-Score using a dataset from the academic postgraduate domain. The results highlight variability across configurations and emphasize how properly designing the processing of retrieved context can improve the quality of generated answers.
This work introduces an innovative AI-driven pipeline for automated 3D CT assessment of post-COVID lung damage, addressing critical limitations of conventional methods. The framework combines a dual-encoder 3D U-Net (Dice = 0.983, 3.5
This work addresses the problem of automatic image classification, focusing on its application to historical press digitization. Specifically, it proposes a method for automatic image discrimination to detect issues such as skew, noise, curvature, and the combination of several previous problems. Empirically, we have observed that addressing some of these image problems increases the quality of optical character recognition and segmentation of newspaper columns. Therefore, it is necessary to predict the problem of a newspaper page and treat it appropriately within the workflow. For this purpose, the problem has been studied by evaluating different pre-trained Convolutional Neural Networks for computer vision problems, such as RestNet variants, AlexNet, VGG11, and EfficientNet. The main contribution of this work is the development of an algorithm for threshold estimation in deep neural networks using metaheuristic optimization, specifically designed for multi-label classification tasks. We employ Particle Swarm Optimization and Genetic Algorithm as representative metaheuristic approaches for optimizing decision thresholds. In addition, we introduce a novel dataset of historical press images, collaboratively annotated by human experts. As a baseline, we propose a Random Search strategy for threshold selection, which is compared against the metaheuristic-based methods. The paper outlines the full computational methodology, including problem formulation, neural architecture design, and dataset construction, and presents experimental results across various deep learning models. The proposed approach achieves consistent improvements in F1-score relative to both the naive thresholding strategy and the baseline method.
Accurate and efficient spacecraft detection is critical for mitigating collision risks in Earth's increasingly crowded orbital environment, where the proliferation of space debris threatens to escalate the Kessler Syndrome a cascading cycle of collisions that generate catastrophic debris fields. This study presents a YOLOv8n-based preliminary framework optimized for real-time spacecraft detection in synthetic images derived from NASA's Pose Bowl Challenge dataset. By performing rigorous preprocessing including bounding box area filtering (75th percentile, ≤ 2500 px2) and resolution reduction from HR to 256 × 256 pixels, we mitigate noise from structural outliers and maintain computational efficiency. Transfer learning on YOLOv8n achieved a validation mAP50–95 of 0.73 at 125 FPS, demonstrating near real-time capability critical for timely collision avoidance, even with limited hyperparameter tuning during 5 epochs and 10 iterations. Training stopped at epoch 356 for a patience value of 25, with peak performance at epoch 212 achieving 0.85 and 1.11 for mAP50 and box loss respectively. While domain gaps from synthetic data and suboptimal convergence highlight the need for deeper architectural refinement, this work underscores the potential of lightweight detection systems to enhance space situational awareness. By enabling rapid identification of spacecraft and debris, the proposed preliminary approach could reduce collision risks, mitigate debris generation, and contribute to long-term orbital sustainability.
Increasing diversity in decision forests without compromising efficiency represents a challenge in machine learning. In this sense, the Proactive Forest algorithm has been proposed as an improvement of Random Forest, by introducing an adaptive scheme for attribute selection based on probabilities. However, the proactive scheme that uses this algorithm depends on the importance of the attributes to build new trees has a disadvantage. At a certain point, the probability of training trees with attributes of low predictive power can increase considerably, affecting the efficiency of the model. As a solution to this possible problem, a new variant is presented, called Proactive Frequency Forest, which incorporates two main components. These components focus on: (1) an initial weighted assignment of attribute selection probabilities, calculated using four statistical techniques, and (2) a dynamic updating strategy based on the frequency of occurrence of attributes during forest construction. Experimental validation, performed on multiple data sets, shows that the proposal significantly increases model diversity without negatively affecting predictive performance.
Water losses due to pipeline breakages are a global issue. Timely detection of these events would help mitigate the waste of this natural resource. Various approaches have been employed to address this problem, ranging from classical Machine Learning algorithms and Digital Signal Processing to Deep Learning architectures. This article presents a Convolutional Neural Network for classifying pressure signals identified as background noise or presence of a burst. The model’s architecture comprises an initial normalization layer, four convolutional layers followed by max pooling layers with ReLU activation functions, a global pooling layer, and a final output neuron with a sigmoid activation function. With a lightweight design intended for implementation on low-power microcontrollers and an end-to-end architecture, it achieves 100
Evaluating mitigating and aggravating circumstances in criminal liability poses significant challenges for judges, who must often make these assessments under uncertainty. While the M-LAMAC model assists judges in this task, its current framework lacks the ability to aggregate intensity values additively when circumstances of the same type must be combined—instead relying on minimum, maximum, or average values. This limitation restricts M-LAMAC’s generality, particularly in cases requiring the application of Articles 81.1 and 81.2 of the Cuban Penal Code. To address this gap, we propose the 2-Tuple Linguistic Bounded Sum (2TLBS), a novel aggregation operator designed to compute the collective intensity of circumstances within the M-LAMAC model. We formally demonstrate its properties (commutativity, associativity, and non-decreasing monotonicity) and validate its practical utility through two criminal case studies. The results show that 2TLBS significantly improve the sanction intervals recommendation, highlighting its potential to enhance judicial decision-making.
Brain tractography, derived from diffusion magnetic resonance imaging (dMRI), has become a key tool for studying the structural connectivity of the human brain. This technique infers the pathways of white matter fibers, generating detailed maps that have significantly impacted both neuroscience research and clinical applications. However, technical limitations still hinder its accuracy and widespread clinical adoption, particularly in surgical planning and the diagnosis of neurological disorders. Major challenges include sensitivity to motion artifacts, lack of standardized validation methods, and difficulties in resolving complex fiber configurations such as crossings and bifurcations. This paper presents a critical review of the current state of tractography, along with a novel methodological proposal based on a microstructural convex optimization model. This approach aims to enhance resolution in anatomically complex regions and reduce false positives in fiber reconstruction. Its potential to improve the reliability of tractography and facilitate its effective integration into clinical practice is also discussed.
Stellar classification, a fundamental aspect of astronomy, offers a structured approach to comprehend and characterize the vast diversity of celestial entities. Here we present a new fine-tuned deep convolutional neural network of 1D separable convolutional blocks for stellar classification based on spectral properties using SSDS-17 data from the Sloan Digital Sky Survey, where class imbalance is evaluated using the MIN class and SMOTE balancing techniques. The results obtained during the performance evaluation confirmed the reliability of the proposed architecture of StellarNet in multi-class stellar classification, achieving remarkable values of about 97
