KPI Science News

LINEAR-ACCURACY ONE-BULLET SILENT DUEL WITH PROGRESSING-BY-ONE-THIRD SHOOTING MOMENTS

Vadim Romanuke — Mon, 29 Dec 2025 00:00:00 +0200

Background. A finite zero-sum game is considered, which models competitive interaction between two subjects. The subject, referred to as the duelist, must take an action (or, metaphorically, shoot the single bullet) during a standardized time span, where the bullet can be shot at only specified time moments. The duelist benefits from shooting as late as possible, but only when the duelist shoots first.

Objective. The objective is to determine optimal behavior of the duelists for a pattern of the duel discrete progression, by which the tension builds up as the duel end approaches and there are more possibilities to shoot.

Methods. Both the duelists act within the same conditions, and so the one-bullet silent duel is symmetric. Therefore, its optimal value is 0 and the duelists have the same optimal strategies. The shooting accuracy is linear being determined by an accuracy proportionality factor.

Results. Depending on the factor, all pure strategy solutions are found for such duels, whose possible-shooting moments comprise a progression pattern. According to this pattern, every next possible-shooting moment is obtained by adding the third of the remaining span to the current moment. The solutions for this pattern are compared to the known solutions for the geometrical-progression pattern and the pattern whose possible-shooting moments progress in a smoother manner.

Conclusions. The proved assertions contribute another specificity of the progressing-by-one-third shooting moments in linear-accuracy one-bullet silent duels to the games of timing. Compared to duels for other duel discrete progression patterns, the specificity consists in that the duel with progressing-by-one-third shooting moments has a constant interval of lower (weaker) shooting accuracies, at which the duelist possesses an optimal pure strategy. This interval is that symmetrically breaks the low-accuracy interval .

DECISION MAKING IN ANTI-CORONAVIRUS DRUG DISCOVERY: MATHEMATICAL MODELING AND VALUE OF INFORMATION ANALYSIS

Dmytro Horodetskyi, Mykhailo Smetiukh, Serhii Soloviov — Mon, 29 Dec 2025 00:00:00 +0200

Background. The process of preclinical evaluation of antiviral medications often consists of multiple stages, each containing substantial uncertainties. Traditional methods for screening the compounds often lack structured means for optimizing the decision-making and calculating the feasibility and risks of transitions between all of the stages. Thus, there appears to be a problem with the inefficient selection of promising antiviral molecules, which subsequently increases the probability of choosing suboptimal research trajectories.

Objective. To develop a computational framework for optimizing of the transition between stages in preclinical antiviral testing. The system focuses on the integration of decision trees and Markov models in order to include effectiveness, risks and the value of additional information into assessment, supporting an in-depth planning of preclinical research pipelines.

Methods. Experimental data from molecular docking, cytotoxicity CD50, and antiviral activity IC50 were used in a multi-stage evaluation system with CTI ≥ 4 being the criterion for progression into further stages. Decision trees provided the explicit rules for advancement of the compounds, while Markov models added context for building sequential strategies under uncertainty and quantified the feasibility of movement to the next stage. Value of information analysis added the assessment of the expected benefit of additional data.

Results. The developed framework consistently produced reliable technical results. The decision used in CTI ≥ 4.0 prediction stage demonstrated a conservative classification pattern, correctly identifying compounds with high therapeutic potential while missing some effective candidates. The Markov model showed steadily increasing state values in docking, cytotoxicity, and antiviral testing phases that confirmed the growth of expected utility. Value of information analysis highlighted that the largest gain occurred after antiviral activity testing.

Conclusions. The study showed that both decision trees and Markov models capture different but complementary aspects of the preclinical evaluation process. Decision trees provide interpretable set of rules that formalize how molecular docking and cytotoxicity measurement influence the progression of compounds, while their limited sensitivity at the CTI threshold highlighted the complexity of prediction the final success of the evaluated compounds. The Markov model simulations showed that the full three-stage pipeline is justified and that progression decisions are influenced by both uncertainty and experimental cost. The value of information analysis clarify the importance of each stage, helping to emphasize the role of antiviral activity data in reducing uncertainty. These findings support the integration of analytic methods for improving the structure, transparency and efficiency of antiviral preclinical research.

A METHOD FOR FRACTAL-DRIVEN REGULARIZATION OF AUTOENCODERS IN SEMI-SUPERVISED MEDICAL IMAGE CLASSIFICATION

Oleksii Zarytskyi, Valerii Danilov — Mon, 29 Dec 2025 00:00:00 +0200

Background: Medical image classification using deep learning is a critical task, yet its effectiveness is constrained by the scarcity of labeled data, which is expensive to acquire. Semi-supervised learning (SSL) methods address this by leveraging unlabeled data. Common autoencoder (AE)-based approaches use reconstruction as a training signal. However, standard reconstruction loss minimization does not guarantee that the resulting latent space will be optimally structured for the classification task, as the model may focus on diagnostically irrelevant features.

Objective: To develop and experimentally validate a novel latent space regularization method: fractal-driven regularization (FDR). The goal is to improve classification metrics for medical images under conditions of severe labeled data scarcity (5%) by integrating fractal dimension (FD) as an additional, a priori training signal.

Methods: The proposed model (FDR-AE) is based on an autoencoder architecture, augmented with two heads attached to the latent space: a classification head and a regression head. The regression head is trained to predict the input image's FD, which is pre-calculated using the "box-counting" method. The total loss function is a combination of three components: classification loss (on 5% labeled data) and both reconstruction and fractal regression losses (on 100% of data). The method's efficacy was validated on three datasets of different modalities (ISIC2024, COVID-19 Radiology, Brain Tumor MRI), comparing it against a baseline convolutional network (Base-CNN) and a standard semi-supervised autoencoder (SSL-AE).

Results: The experiments demonstrated a consistent advantage for the proposed method. On the ISIC2024 dataset, FDR-AE achieved an F1-Score of 0.508 for the "malignant" class, compared to 0.431 for SSL-AE and 0.304 for Base-CNN. On the COVID-19 dataset, the F1-Score for the "covid19" class was 0.722 for FDR-AE versus 0.695 for SSL-AE. In the 4-class Brain Tumor task, FDR-AE showed improved F1-Scores across all classes, with the most significant gains (+0.079 and +0.054) observed for classes 0 and 3, which also had the greatest mutual statistical difference in their FD values.

Conclusions: Fractal-driven regularization demonstrates that FD is a valuable a priori signal for learning higher-quality, structurally-grounded representations in SSL tasks. The method is particularly effective on simple architectures under severe data scarcity. Prospects for future research include using FDR as a pre-training method or implementing a dynamic coefficient for the regression component of the loss function.

DEEP LEARNING-BASED MELANOMA CLASSIFICATION ENHANCED BY FRACTAL DIMENSION ANALYSIS

Vladyslav Nikitin, Valerii Danilov — Mon, 29 Dec 2025 00:00:00 +0200

Background. Melanoma is a malignant skin lesion that is prone to metastasize aggressively, leading to an almost guaranteed lethal outcome if left unchecked. In contrast, early-stage detection allows for the tumor to be removed via a harmless surgical procedure that may not even leave a scar. However, the availability of competent diagnostics are often limited due to a shortage of healthcare specialists and technologies. Deep Learning models such as Visual Transformer (ViT) have demonstrated strong performance, but researchers continuously seek to improve the results by incorporating new features. Since human skin exhibits fractal-like characteristics, it is theorized that metrics quantifying this complexity can act as valuable supplementary features for DL models, leading to increased classification accuracy.

Objective. We investigated the impact of the integration of fractal dimension (FD) on a Vision Transformer deep learning model used for melanoma classification. A comparison was made on models that received random noise vs. the estimation of FD value.

Methods. Vision Transformer was used as a feature-extracting backbone pre-trained on ImageNet dataset. Fine-tuning was done on this backbone in combination with a classification head targeted to distinguish melanoma vs. nevus classes. Along with extracted features, the classification head received FD value. An identical model received random noise instead of FD. Statistical testing and FD impact analysis were done to confirm the significance of the new feature.

Results. Integrating FD into ViT showed noticeable improvement in test metrics. SHAP analysis confirmed the meaningfulness of the new feature. McNemar's test validated that the difference in model predictions was statistically significant.

Conclusions. The results suggest that FD can serve as a valuable supplementary feature for DL models, and the integration of biomarkers such as FD provides a basis for more robust melanoma classification.

SYSTEM APPROACH TO MULTICRITERIA EVALUATION OF SESSION-BASED AND SEQUENTIAL RECOMMENDATION SYSTEMS

Nadezhda Nedashkovskaya, Dmytro Androsov — Mon, 29 Dec 2025 00:00:00 +0200

Background. Recommendation systems have become indispensable components of modern digital platforms, enabling personalized content delivery across diverse domains. Traditional collaborative filtering and content-based approaches often fail to capture temporal dynamics and contextual dependencies inherent in user behavior patterns. Sequential recommendation systems (SRSs) and session-based recommendation systems (SBRSs) have emerged as new paradigms to capture users' short-term but dynamic preferences for enabling more timely and accurate recommendations.

Objective. To propose a system approach for multicriteria evaluation of various SRS and SBRS models – a unified framework for understanding these models, selecting the best recommendation model, and guiding future research directions in temporal-aware recommendation systems. To provide a systematic overview and comprehensive analysis of session-based and sequential recommendation systems, to examine their theoretical foundations, evolution, empirical performance characteristics, and practical deployment considerations.

Methods. A comprehensive analysis of foundational approaches from Markov chain models to modern neural architectures including attention-based methods, graph neural networks, and state-space models is conducted. The approaches are systematically categorized based on architectural principles, temporal modeling strategies, and knowledge integration methods. The Analytic Hierarchy Process is applied for calculation of relative importance of benefits, costs, opportunities and risks in a problem of session-based and sequential recommendation systems synthesis. An experimental study of various SRS and SBRS models was performed on benchmark datasets.

Results. Empirical studies on temporal benchmark datasets show that combining SASRec and ReCODE improves the Recall@K metric by 9% over the baseline SASRec model, and combining GRU4Rec with ReCODE improves the metric by 17% over the baseline GRU4Rec. The SASRec model, which adapts transformer architectures to the sequential recommendation problem, achieved the highest baseline performance in terms of Recall@K and NDCG@K criteria on benchmark datasets compared to the other examined models, demonstrating the effectiveness of self-attention mechanisms for sequence modeling. ReCODE is a model-independent neural ordinary differential equation framework for recommender systems and an effective framework for studying consumer demand dynamics, has improved the metrics of existing baseline approaches, and has acceptable computational complexity for practical recommender system deployment scenarios.

Conclusions. Session-based and sequential recommendation systems have evolved through several paradigmatic shifts with significant scientific achievements including establishment of session-based recommendation model as distinct from traditional collaborative filtering, development of attention mechanisms for sequence modeling, and introduction of continuous-time formulations. Future research directions include unified architectures, scalability solutions, improved evaluation methodologies, and extensions to multi-stakeholder scenarios.

A ARCHITECTURE OF CNN-TRANSFORMER HYBRID WITH MASKED TIME SERIES AUTO-CODING FOR BEHAVIORAL BIOMETRICS ON MOBILE DEVICES

Mariia Havrylovych — Mon, 29 Dec 2025 00:00:00 +0200

Background. Continuous behavioral authentication (keystroke dynamics, touch/swipe, motion sensors) verifies identity without extra actions. However, models degrade under device, session and activity shifts, are sensitive to noise and often require significant labeling. As passwordless logins spread, demand rises for post-login risk control and for models that are robust, compute-efficient and stable in the wild.

Objective. To develop and empirically study a compact CNN-Transformer hybrid with lightweight self-supervised masked time-series autoencoding (MAE-style) for mobile behavioral biometrics on the HMOG and WISDM datasets.

Methods. A 1D-CNN front end extracts local cues from smartphone motion signals, while a Transformer encoder captures longer-range dependencies. We use masked reconstruction on unlabeled HMOG sessions for self-supervised pretraining under a limited computational budget and then fine-tune the same hybrid architecture for user identification. We evaluate three hybrid variants on HMOG (trained from scratch, with masked pretraining, and with masked pretraining plus CORAL domain adaptation) and three models on WISDM (a Transformer baseline, a hybrid trained from scratch and a hybrid initialized from the HMOG-pretrained weights). Performance is measured using user-level mean and median Equal Error Rate (EER) and AUC.

Results. On HMOG, the hybrid model trained from scratch achieves the best user-level metrics (EER 21.51% mean, 18.63% median; AUC 0.854 mean, 0.905 median), while the lightweight MAE and CORAL variants do not yet surpass this baseline. On WISDM, the hybrid model substantially outperforms a pure Transformer baseline (EER 9.41% vs 51.25% mean; AUC 0.902 vs 0.488 mean), and cross-dataset initialization from the HMOG MAE-pretrained weights provides an additional improvement (EER 8.42% mean, 2.07% median; AUC 0.907 mean, 0.959 median).

Conclusions. The results indicate that a compact CNN-Transformer hybrid is effective for sensor-based mobile behavioral biometrics and that even lightweight masked pretraining can be helpful for cross-dataset transfer. At the same time, the benefits of MAE and CORAL on HMOG depend strongly on the pretraining budget and masking configuration, suggesting that further tuning is needed to fully exploit self-supervised pretraining in this setting.