System research and information technologies

Analysis of actuarial risk with generalized linear models

Roman Panibratov, Petro Bidyuk — Mon, 29 Dec 2025 00:00:00 +0200

The problem of applying generalized linear models to the analysis of actuarial risks in the context of premium charges to clients was considered. The Monte-Carlo method for Markov chains was applied. Two situations were considered for the computational experiment. For the first one, insurance indicators and the target variable were randomly assigned due to the problem of public data access. To create three datasets, charges were generated from normal, gamma, and Pareto distributions with dynamic variance, and noise was added to stimulate a non-stationary process. In the second situation, actual actuarial data from the Singa-pore Actuarial Society was used. Generalized Linear Models with normal dis-tribution and logarithmic link function, an exponential distribution and loga-rithmic link function, and Laplace distribution with identity link function were constructed. Based on the model-fitting quality metrics, conclusions were drawn about their structure.

The results of the multi-position surveillance system’s efficiency, depending on the locations of its sensors, using additional data processing

Mon, 29 Dec 2025 00:00:00 +0200

The efficiency of a multi-position system depends on the realization of its struc-ture — how many elements it includes, where they are located, and how the envi-ronment and terrain influence its operation. The paper is dedicated to data pro-cessing in a multi-position surveillance system as an additional option, leverag-ing the in-between big data from the system’s elements. A sufficient number of numerical data generated by the multi-position system and its elements–sensors–allows the use of statistical methods and models from machine learning or deep learning. The ontology for quality estimation of the multi-position sys-tem, depending on its configurations, is proposed. The results of the distribu-tions of detected events are presented in graphical forms that allow statistical evaluation of the distributed data. Our findings allow us to ensure the efficiency of a multi-position system in an unpredictable, variable environment by recon-figuring it when it offers better capabilities.

Matrix-graphic simulation of social network: ergodic properties

Igor Spectorsky, Vitalii Statkevych, Oleksandr Stus — Mon, 29 Dec 2025 00:00:00 +0200

We propose mathematical tools for social network simulation to obtain sufficient conditions for network ergodicity, defined as the existence of a steady state as time approaches infinity. The proposed model is linear; the network elements form a two-dimensional array (matrix), where each entry represents the state of an element at a specific time.

An impact operator, structured as a four-dimensional array, defines the interactions between elements. This operator is also presented as a directed graph where vertices correspond to network elements, and arcs represent the impact of one element on another. The model incorporates boundary elements that influence the internal states of the network.

Sufficient conditions for network ergodicity are derived from the connectivity properties of the impact graph, which must contain paths between all pairs of vertices and loops for all vertices. These conditions ensure that the operator's spectrum (with the possible exception of the value 1) is located inside the open unit disk. We prove that 1 is an eigenvalue if and only if the boundary is isolated. These spectral properties guarantee that a steady state exists and can be found using an iterative procedure with linear (geometric) convergence.

Overview of neural network object detection methods & models on the example of their use for lab animal observation

Maksym Shvandt, Volodymyr Moroz — Mon, 29 Dec 2025 00:00:00 +0200

This article provides a brief overview of a set of the most common basic object detection neural network models. Today, the need for automating surveillance and observation processes remains a growing trend. Moreover, one of the key tasks of such processes is usually the detection of an object of interest for further analysis. Previously, many basic object detection algorithms and approaches have been proposed; however, most of them typically have limitations in terms of their applicability. In most cases, these limitations arise due to the nature of the observed environment or because the detection approaches rely on specific object characteristics, such as color or basic shapes only. To address these problems, a new approach for object detection has been developed using neural networks. This paper presents the basis and central aspects of the most common neural network object detection models. The experiment has demonstrated the features, advantages, and disadvantages of the studied methods in the application case of lab animal detection during their behavioral study. Considering this, conclusions and recommendations for their usage cases were made.

Quality assessment of models and deep learning methods for super-resolution image formation

Anna Lanko, Nadezhda Nedashkovskaya — Mon, 29 Dec 2025 00:00:00 +0200

This article examines evaluation metrics for the results of super-resolution image generation in solving the SISR task. The study comprises two experiments: the implementation of custom network architectures for SRGAN, VDSR, and SRCNN, and fine-tuning of pre-trained SRGAN, VDSR, and SRCNN models. An algorithm for assessing the quality of models and deep learning methods for generating super-resolution images is suggested. The VDSR model performed best in terms of pixel, structural, and perceptual metrics, as well as training time and visual confirmation by a human, highlighting that residual learning is more effective than recursive learning under the conditions of the two conducted experiments. Threshold values for practically acceptable and high-quality results were determined through visual analysis of many generated images and their corresponding quality metrics, including those reported by other researchers.

Prediction of mechanisms of toxic action of phenols by means of probabilistic neural network in combination with Kruskal–Wallis test

Yaroslava Pushkarova, Galina Zaitseva — Mon, 29 Dec 2025 00:00:00 +0200

Prediction of the toxicity of chemical compounds is one of the most important steps in drug design. The use of phenolic compounds is a promising component in the pharmaceutical industry with many possible applications. The paper fo-cuses on the application of a probabilistic neural network for classifying 232 phenols based on their mechanisms of toxic action. The Kruskal–Wallis test was also used to assess the influence of molecular descriptors on the reliable classification of phenolic compounds based on the mechanisms of their toxic action. It is shown that for the correct training of a probabilistic neural network and effective prediction of the mechanisms of toxic action of phenols, it is suf-ficient to use only 5 molecular descriptors.

Development of algorithms for detecting defects in the code sequence structure on the surface of modulation disks

Mon, 29 Dec 2025 00:00:00 +0200

This study investigates algorithms for detecting and localizing defects in code sequence structures on modulation disk surfaces. It targets small anomalies in lithographically patterned elements that can cause readout errors or reduced measurement accuracy. A multi-level image-processing model combines Gaussi-an smoothing, adaptive thresholding, morphological operations, and contour-based segmentation. Processing stages are formalized as mathematical operators for reproducible implementation. Defects are characterized using perimeter- and area-based metrics, and their area distribution is approximated by a normal law. A spatial model computes defect centroids, enabling comparative quality as-sessment of disk samples. The software provides an interface for tuning thresh-olds, visualizing contours and defect-area plots, and exporting results. Tests on real defective disks confirm the method’s reliable detection of local structural violations and its suitability for diagnostic systems.

Algorithms for assignment of external reviewers for PhD-thesis defense

Serhiy Shtovba, Mykola Petrychko — Mon, 29 Dec 2025 00:00:00 +0200

We propose an approach to assigning external reviewers. In the proposed ap-proach, only the semantic similarity between applications and reviewers is tak-en into account; the similarity indices are assessed, and the necessary number of reviewers is assigned to ensure the maximum suitability level of the reviewers with the application, according to some criteria. We also perform a comparative analysis of various optimization algorithms using the criterion of “assignment quality–optimization time”. Experiments on the dataset showed that a reasona-ble balance between the “assignment quality” and “optimization time” criteria for the assignment of external reviewers can be achieved using a greedy algo-rithm without elitism or brute-force search on a truncated set of candidates. An application of the proposed algorithms improves the average quality of PhD committees by 13–34% across the entire dataset, depending on the algorithm used.