Publicações

Random number generators are extensively used in science. Generating pseudo-random numbers is the base for many data analysis techniques in computational statistics. This is the case, for instance, of most of the Bayesian methods, which are enabled by means of samplers such as the well-known Gibbs sampler and the Metropolis–Hastings. These classical Markov Chain Monte Carlo samplers are designed to generate a sequence of numbers that, under certain conditions, converge to a sequence that behaves as if sampled from a user-defined target distribution. In general, the number of iterations required to reach such convergence is not deterministic. There are several statistical tests for identifying that convergence has not yet been achieved, but not for actually signaling convergence. The present work introduces an exact non-parametric sequential test for signaling the convergence of random number generators in general. The solution is derived in the light of the type I error probability spending approach.

Multi-objective optimization problems (MOPs) often require a trade-off between conflicting objectives, maximizing diversity and convergence in the objective space. This study presents an approach to improve the quality of MOP solutions by optimizing the dispersion in the decision space and the convergence in a specific region of the objective space. Our approach defines a Region of Interest (ROI) based on a cone representing the decision maker's preferences in the objective space, while enhancing the dispersion of solutions in the decision space using a uniformity measure. Combining solution concentration in the objective space with dispersion in the decision space intensifies the search for Pareto-optimal solutions while increasing solution diversity. When combined, these characteristics improve the quality of solutions and avoid the bias caused by clustering solutions in a specific region of the decision space. Preliminary experiments suggest that this method enhances multi-objective optimization by generating solutions that effectively balance dispersion and concentration, thereby mitigating bias in the decision space.

Deep Learning (DL) methods offer competitive alternatives for Customer Churn Prediction (CPP) but are challenging to optimize due to their complexity. This study introduces STArS (Sequential Temporal Neural Architecture Search), a Genetic Algorithm-based Neural Architecture Search (NAS) method, to optimize Temporal Convolutional Neural Networks (TCNN) and Long Short-Term Memory (LSTM) models. The optimized STArS models outperform traditional models, such as Random Forest and XGBoost, as well as an empirically developed TCNN. The research also highlights the importance of feature analysis for retention strategies, using a real dataset from a financial services provider. This demonstrates the practical applicability of NAS techniques across multivariate time series datasets.

Studies on spatial cluster patterns are of interest in many areas. Spatial scan statistics is the most widespread strategy for studying these patterns. However, scan statistics lose substantial efficiency in situations where candidate clusters can assume irregular shapes. Conversely, other techniques, with the aim of increasing the flexibility of analyzing cluster shapes, have emerged. We present two novel reinforcement learning approaches that use scan spatial statistics to represent the reward function. The novel approaches are explained in detail, and there is an extensive set of computational experiments with controlled synthetic data to verify their functionality and adaptation to the problem of detecting spatial clusters. Our results attest to the quality and applicability of the new techniques for addressing this problem.

Community detection in complex networks is fundamental across social, biological, and technological domains. While traditional single-objective methods like Louvain and Leiden are computationally efficient, they suffer from resolution bias and structural degeneracy. Multi-objective evolutionary algorithms (MOEAs) address these limitations by simultaneously optimizing conflicting structural criteria, however, their high computational costs have historically limited their application to small networks. We present HP-MOCD, a High-Performance Evolutionary Multiobjective Community Detection Algorithm built on Non-dominated Sorting Genetic Algorithm II (NSGA-II), which overcomes these barriers through topology-aware genetic operators, full parallelization, and bit-level optimizations–achieving theoretical \$\$O(G \backslashcdot N\_p |V|)\$\$complexity. We conduct experiments on both synthetic and real-world networks. Results demonstrate strong scalability, with HP-MOCD processing networks of over 1,000,000 nodes while maintaining high quality across varying noise levels. It outperforms other MOEAs by more than 531 times in runtime on synthetic datasets, achieving runtimes as low as 57 s for graphs with 40,000 nodes on moderately powered hardware. Across 14 real-world networks, HP-MOCD was the only MOEA capable of processing the six largest datasets within reasonable time, with results competitive with single-objective approaches. Unlike single-solution methods, HP-MOCD produces a Pareto Front, enabling individual-specific trade-offs and providing decision-makers with a spectrum of high-quality community structures. It introduces the first open-source Python MOEA library compatible with networkx and igraph for large-scale community detection.

Arrhythmia, a cardiac condition, is frequently diagnosed by classifying heartbeats using electrocardiograms (ECG). This classification is a crucial step in medical diagnosis and can be significantly improved by employing computational methods to analyze the ECG data. Despite the extensive literature on this subject, the high inter-patient variability and noise in ECG signals pose challenges in the development of computational methods. Deep learning methods represent state-of-the-art solutions to diverse problems in computer vision, signal processing, and pattern recognition, mainly due to advancements enabled by self-supervised learning. In this work, we propose a self-supervised approach for ECG beat classification and a specific pretext task for ECG, termed ECGPuzzle. This approach allows for fine-tuning a deep learning model to an individual, improving the model's generalization. Given the low computational power of medical equipment and the need for on-site training, hardware acceleration is indispensable. Thus, we investigate the feasibility of this proposal on three distinct computational systems and discuss potential manners to train a model on an embedded system.

Accurately identifying plant species and varieties is crucial across various disciplines, such as biology, medicine, and agronomy. While species identification is challenging, variety identification presents an even greater difficulty. Conventional identification methods, although effective, often require specialized and costly equipment, making them less accessible. In this work, we explore the problem of hop variety classification, comparing traditional feature extraction methods with deep learning approaches using the UFOP-HVD dataset. We address two research questions: whether traditional techniques can achieve competitive results given the limited number of images and whether combining traditional techniques and deep learning can improve the current state-of-the-art. Our findings indicate that traditional techniques yield competitive results for hop variety identification, offering advantages such as interpretability, reduced computational costs, and potential integration into mobile devices. Moreover, we introduce an ensemble method that improves the accuracy from 77.16% to 81.90%, establishing a new state-of-the-art for the UFOP-HVD dataset. These results demonstrate the potential of merging traditional methods with deep learning for challenging hop variety classification tasks, providing an initial baseline for future research.

The accurate identification of promoter regions in DNA sequences holds significant importance in the field of bioinformatics. While this problem has garnered substantial attention in the literature, it remains unresolved. Several researchers have achieved notable outcomes by employing diverse machine-learning techniques to predict promoter regions. However, only a few have thoroughly explored the utilization of features derived from the physicochemical properties of DNA across various organism types. This study investigates the advantages of incorporating these features in the training of machine-learning models. The research evaluates and compares the performance of multiple metrics on diverse datasets encompassing both prokaryotic and eukaryotic organisms. The state-of-the-art CNNProm method is employed as the baseline for our experiments. The models and source code associated with this study can be accessed at the following URL of the project's repository: https://anonymous.4open.science/r/bracis-paper-1458/.

In this study, we examine the impact of human mobility on the transmission of COVID-19, a highly contagious disease that has rapidly spread worldwide. To investigate this, we construct a mobility network that captures movement patterns between Brazilian cities and integrate it with time series data of COVID-19 infection records. Our approach considers the interplay between people's movements and the spread of the virus. We employ two neural networks based on Graph Convolutional Network (GCN), which leverage spatial and temporal data inputs, to predict time series at each city while accounting for the influence of neighboring cities. In comparison, we evaluate LSTM and Prophet models that do not capture time series dependencies. By utilizing RMSE (Root Mean Square Error), we quantify the discrepancy between the actual number of COVID-19 cases and the predicted number of cases by the model among the models. Prophet achieves the best average RMSE of 482.95 with a minimum of 1.49, while LSTM performs the least despite having a low minimum RMSE. The GCRN and GCLSTM models exhibit mean RMSE error values of 3059.5 and 3583.88, respectively, with the lowest standard deviation values for RMSE errors at 500.39 and 452.59. Although the Prophet model demonstrates superior performance, its maximum RMSE value of 52,058.21 is ten times higher than the highest value observed in the Graph Convolutional Networks (GCNs) models. Based on our findings, we conclude that GCNs models yield more stable results compared to the evaluated models.