Mariá Cristina Vasconcelos Nascimento

Time series anomaly detection (TSAD) is crucial for identifying unusual patterns in sequential data across domains such as finance, healthcare, cybersecurity, and predictive maintenance. A key challenge is selecting the most suitable anomaly detection algorithm for a given time series, due to the wide variability in instance characteristics. This paper tackles the algorithm selection problem for univariate TSAD by applying meta-learning to predict the performance ranking of candidate methods. We propose a meta-learning framework that extracts statistical, structural, and temporal meta-features to model algorithm behavior. The study evaluates four TSAD algorithms-two model-based (DeepAnt and Temporal Convolutional Autoencoder) and two discord detection methods (Matrix Profile and Merlin). Experimental results show that meta-learning is effective for unsupervised TSAD algorithm selection. A Random Forest meta-model achieved strong results, with a Spearman correlation of 0.61, Kendall tau of 0.52, top-1 accuracy of 64%, and top-2 accuracy of 89%, demonstrating its reliability even when performance differences between algorithms are small.

The formation of surgical kits is essential for efficient resource management in hospitals. This problem involves determining the appropriate items and quantities to ensure that kits meet the demands of surgical procedures without shortages. A key challenge is designing shared kits that fulfill overlapping demands across procedures without including excess items, which may be discarded once deemed contaminated. Unlike existing approaches that allow oversupply, we propose optimization-based methods, including a mixed-integer linear programming model (MILP) and a two-stage heuristic, to minimize surplus in shared kits. The first stage applies a biclique enumeration framework from bipartite graph clustering. Computational experiments using real data from a Brazilian hospital demonstrate that the proposed methods effectively identify shared kit compositions. Although the heuristic yielded lower-quality solutions than the solver for larger instances, it successfully uncovered meaningful patterns that offer valuable insights andwarrant further exploration. In addition, tests conducted with a commercial solver using the proposed mathematical model and the heuristic results as initial solutions revealed alternative solutions with a higher number of shared kits in certain instances.

This paper addresses the problem of managing perishable inventory under multiple sources of uncertainty, including stochastic demand, unreliable supplier fulfillment, and probabilistic product shelf life. We develop a discrete-event simulation environment to compare two optimization strategies for this multi-item, multi-supplier problem. The first strategy optimizes uniform classic policies (e.g., Constant Order and Base Stock) by tuning their parameters globally, complemented by a direct search to select the best-fitting suppliers for the integrated problem. The second approach is a hyper-heuristic approach, driven by metaheuristics such as a Genetic Algorithm (GA) and Particle Swarm Optimization (PSO). This framework constructs a composite policy by automating the selection of the heuristic type, its parameters, and the sourcing suppliers on an item-by-item basis. Computational results from twelve distinct instances demonstrate that the hyper-heuristic framework consistently identifies superior policies, with GA and EGA exhibiting the best overall performance. Our primary contribution is verifying that this item-level policy construction yields significant performance gains over simpler global policies, thereby justifying the associated computational cost.

Predicting the reference signal received power in wireless communication is crucial for improving network performance, allocating resources, and ensuring good signal coverage, especially in advanced technologies like 5G and beyond. To create an accurate prediction model, we need to look at different aspects of the radio environment and understand the importance of each factor. In our study, we suggest using machine learning to predict reference signal received power. We analyze the importance of features by studying their impact on the received signal power. We developed a machine learning approach using 64 features taken from recent literature and new ones proposed in this study from real-world received signal power measurements in outdoor areas, including cities and suburbs. Using this data, we trained a Random Forest model for received signal power predictions. After training, we analyzed the importance of each feature to create a simpler machine learning model that maintains good prediction accuracy. Our results show that we can use only the 25 most important features to build a less complex model with a small error difference of 0.14 dB compared to the original model with 64 features.