Stroke segmentation is crucial for interpreting free-hand sketches, supporting tasks ranging from recognition and generative modeling to semantic analysis. However, existing methods often struggle to distinguish between similar component categories and typically rely on large models unsuitable for on-device deployment. In this paper, we present InstanceSketch, a two-step pipeline designed to segment complex sketch objects, and InstanceSketch-Scene, an extension that decomposes entire scenes containing multiple objects into semantically meaningful parts. By integrating convolutional neural networks and Transformers, our approach enables efficient breakdown of sketches into interpretable components. An important advancement is a novel method for refining labels of similar components using clustering algorithms to ensure accurate separation between visually similar stroke groups. Our proposed method not only matches but also exceeds the performance of existing state-ofthe-art techniques, while operating efficiently with significantly fewer resources, making it an attractive solution for on-device applications and serverless environments.
Privacy compliance is a critical requirement for legal entities handling personal data (PD), demanding the integration of protective mechanisms into business workflows and transparency with data subjects (DSs). However, a critical gap persists: DSs struggle to understand privacy information and effectively use available protections. This human-centered challenge demands alignment between organizational processes and individuals' cognitive and behavioral capacities. Privacy heuristics (PHs) can bridge this gap by supporting user decision-making, yet their design is complex, prone to bias, and, if done irresponsibly, may lead to unethical or manipulative outcomes. This paper presents design principles for creating Responsible Privacy Heuristics (RPHs) in usable privacy-aware systems. We demonstrate applicability through online social network examples and validate through an A/B test with 12 users. Results show RPHs matched standard PHs in usability while proving slightly more effective at preventing privacy-invasive choices and fostering informed decision-making, without compromising user autonomy.
In cybersecurity, analysing network traffic is critical for identifying potential threats and mitigating incidents. Traditional approaches to network traffic analysis often rely on synchronous methods that may not fully capture the dynamic nature of network behaviour. This paper presents an approach for asynchronous evidence-based network traffic assessment, experimenting on synthetic cyber-attack templates and large network flow datasets available online. Our approach leverages asynchronous data collection to capture network traffic at varying time intervals, enabling the identification of incidents across different time frames and locations by processing and aligning the data. By combining asynchronous data collection with evidence-based assessment, our approach enables cybersecurity analysts to gain deeper insight into network traffic dynamics, enhance threat detection capabilities, and improve incident response effectiveness. We demonstrate the effectiveness of our approach through experimental evaluations using synthetic cyber-attack templates in a virtual environment, while capturing network flows. In summary, our research advances the field of network traffic analysis by presenting an approach that addresses the shortcomings of conventional synchronous techniques and lays the groundwork for more resilient, adaptive cybersecurity solutions.
Predicting order success in business-to-business environments remains challenging due to the multifactorial complexity of transactions and limitations in traditional forecasting. This study develops a multi-algorithmic method to identify and rank factors influencing order success, with empirical validation in an ERP-based transactional environment. It integrates five feature significance scorers (Area Under the Curve, Mutual Information, distance correlation, Logistic Regression, and Decision Trees) with modified Borda rank aggregation. Analysis of 86,794 orders spanning 2017-2025, with 25 characteristics, revealed that communication factors, particularly message count, demonstrate the strongest influence on success. Historical cooperation experience ranks second, while financial parameters showed ambiguous results, and discounts proved insignificant. Order flexibility positively correlates with success, whereas temporal characteristics showed no influence. The study contributes a methodologically justified five-scorer selection, a modified Borda aggregation method for consolidated scale-independent factor ranking, and a reproducible pipeline for structured transactional data.
Postural sway during quiet standing is often analyzed using diffusion-or Brownian-type models that implicitly assume near-Gaussian statistics. In practice, center-of-pressure (COP) trajectories can exhibit heavy tails, asymmetry, and intermittent excursions that may bias variance-only variability descriptors. This study presents a reproducible Python-based analysis pipeline for assessing non-Gaussian structure in COP increment series by rotating increment pairs into a PCAdecorrelated coordinate system. Using the Human Balance Evaluation Database (HBEDB) (1930 trials from 163 subjects; 60 s at 100 Hz under four standing conditions: eyes open/closed and rigid/unstable surface; three repetitions each), we compare empirical increment distributions with mean-and covariance-matched Gaussian surrogates and quantify departures using complementary diagnostics, including probability-probability comparisons, robust spread descriptors (IQR and range), and uni-/bivariate kernel density estimation. We further introduce a compact central ellipse fraction (CEF) descriptor that measures central concentration in the PCA plane and relates distributional findings to linear and nonlinear variability measures, including Poincar & eacute;-type descriptors and recurrence indices. Across the dataset, empirical increments deviate systematically from Gaussian surrogates: P-P plots show tail departures, and the 2-D density in the PCA plane is more centrally concentrated than the matched Gaussian model (CEF higher by approximately 0.08-0.10), consistent with leptokurtic (peaked and heavy-tailed) behavior. These results motivate distribution-aware preprocessing and descriptor selection for computer-based posturography and fall-risk assessment algorithms beyond variance-only summaries.
This paper introduces the Hybrid Physics-Informed Temporal Attention (H-PITA) model, a novel framework for accurate, physically consistent daily photovoltaic (PV) energy (kWh) forecasting. By integrating a temporal-attention encoder with external physical priors via gated fusion and a lightweight residual pathway, H-PITA balances data-driven learning with domain specific constraints. The model incorporates physical knowledge through symmetric and asymmetric penalties, non-negativity constraints, and dynamic loss weighting, as well as an optional inference-time safety cap. Evaluated on a real-world PV system in Southeast Europe using longitudinal dataset (2014-2025), H-PITA demonstrates superior performance over baselines including LSTM, XGBoost, persistence, climatology, and linear/ridge models, achieving a test R2 of 0.78. Results indicate that data filtering and physical grounding significantly enhance generalization, reducing MAPE and RMSE. Overall, H-PITA provides a robust, modular, and transferable solution for risk-aware PV forecasting, making it highly suitable for operational grid integration and solar asset management.
The escalating complexity of cybersecurity threats requires adopting innovative strategies for the collection and analysis of threat intelligence. This study offers a comprehensive longitudinal analysis of the dissemination of cybersecurity information via Telegram. The research utilizes a multi-phase pipeline that incorporates statistical anomaly detection, machine-learning classification, and sophisticated visualization methods. Over a period of one year, 9,415 messages from nine Telegram channels dedicated to cybersecurity were examined. Collectively, these messages produced more than 50 million interactions. The proposed approach combines Z-score analysis with percentage-based rolling averages to detect anomalies. It has an 80% success rate in linking anomalies to real incidents. The results show clear temporal activity patterns, with 76.9% occurring on weekdays, which corresponds to professional threat intelligence cycles. The January anomaly cluster represents 15.7% of all yearly anomalies. Visual content strategies led to a 3.2x higher engagement rate compared to text-only posts. Additionally, a paradox was observed between content quality and reach: low-volume, specialized channels garnered significantly higher engagement per post. An artificial intelligence-driven classification system using a large language model has categorized messages into 18 distinct threat types, achieving 91.2% accuracy and demonstrating high confidence. These findings contribute to the advancement of cyber threat intelligence derived from social media, offering practical insights for security operations centers, threat hunters, and researchers seeking to leverage social platforms for early warning.
Recent vertical crustal movements in Bulgaria have been mapped using adjusted levelling data from three epochs, yet results from different research teams are inconsistent and often contradict known tectonic fault structures. To provide an independent evaluation, we applied a new estimation approach based on data from Bulgaria's Second (1953-1957) and Third (1975-1980) levelling campaigns. Both networks were processed through 3' independent adjustments, selecting line elevations that minimised loop misclosures. In addition, Inverse Distance Weighting (IDW) with a power parameter of 6 was applied. This procedure reduced benchmark height standard errors to below +/- 5.3 mm and vertical velocity errors to +/- 0.30 mm/year. The resulting high-accuracy map reveals a strong correlation between tectonic boundaries and major earthquake epicentres.
Precise geometric levelling has been the most accurate method for measuring height differences on Earth for over 150 years. Yet, its accuracy has seen little improvement in the past 80 years. This stagnation is mainly due to the rise of GNSS technologies and a lack of progress in levelling data processing. This article presents new statistical insights into how levelling uncertainties accumulate and introduces a revised algorithm for adjusting levelling networks. Testing the method with data from Finland's Third Levelling (1978-2006) under 66% showed standard deviations of benchmark geopotential numbers below 0.7 mgpu (or 0.7 mm). These results prove that, by rethinking outdated approaches, a fifteenfold increase in accuracy is achievable.
The aim of this study is to identify the acoustic similarities and differences among directives. The data consist of specially constructed sentences read by four male actors, representing three subtypes of directive: request, command, and advice (a total of 340 tokens). The following parameters were analysed: mean F0, F0 maximum, F0 range, intonation contour, mean intensity and intensity maximum, as well as the duration of some segments. The results show that the parameters of F0 vary, differ only slightly in some cases, and do not reveal clear tendencies. The intonation contour is falling or rising-falling in the examples containing two intermediate phrases. Intensity is a reliable indicator of active directives - commands and pleas. Phrases expressing advice show the lowest intensity values. Longer duration signals requests.
Personality traits are indicative of consistent patterns in thoughts, emotions, and behaviours, and they consequently impact students' academic performance. Numerous scientific investigations have explored the determinants of student learning success, with some emphasizing the significance of considering both student personality traits and generational cohorts. This study seeks to examine the associations between the Big Five personality dimensions and the learning strategies employed in mathematics lectures. A random sample of first-and second-year informatics engineering science students from Vilnius Gediminas Technical University (VILNIUS TECH) was selected for this research. Data were gathered through an anonymous survey utilizing a questionnaire developed by the authors. Non-parametric statistical tests were employed to analyse the data. The findings revealed that the sample predominantly consisted of students exhibiting the personality dimensions of agreeableness and conscientiousness. Furthermore, students' selection of learning strategies was generally independent of their personality dimensions. Nonetheless, the study underscored students' preferences for diverse learning strategies, irrespective of their personality traits.
This study examines the assessment of human drowsiness using single-channel data from a forehead electrode processed with a spectral analysis algorithm. Spectral band analysis allows for the identification of key parameters related to drowsiness. Eye blink frequency was identified as a useful parameter based on the analysis of signal amplitude and time-frequency characteristics. Data were obtained under two conditions - after 20 hours of wakefulness and after a full night of sleep while participants read an e-book. Spectral analysis calculations and Random Forest and statistical algorithms were used for signal processing to identify the most informative features for the expert decision-making system. The analysis showed a close correlation between spectral indicators (especially alpha and beta bands) and subjective ratings of the Karolinska Sleepiness Scale. Eye blink frequency was also successfully determined using biopotential and video analysis. Expert judgment complements the logical relationships of the parameters for real-time fatigue monitoring, with applications in safety-critical situations and human-computer interaction.
Software Ecosystems (SECOs) represent dynamic and evolving networks of interdependent platforms, components, and services developed through collaborative efforts. Engineering such ecosystems presents unique challenges that traditional software engineering methods are often ill-equipped to handle, particularly given the increasing importance of human actors within modern SECOs. To address this gap, this study introduces a novel, five-stage, human-centered approach for engineering Software Ecosystems (SECOs). Developed by integrating established SECO best practices with empirical insights from the H2020 PHArA-ON (Pilots for Healthy and Active Ageing in Europe) project, our approach is consistent with ISO 9241-210 guidelines. It comprehensively covers the SECO lifecycle-from context-of-use analysis and requirements specification to design, implementation, and continuous evaluation. The approach's relevance and utility were evaluated and affirmed through expert interviews, which also provided valuable feedback for its refinement.
In the modern knowledge era, the exponential growth of digital solutions has led to the generation of vast amounts of data. This necessitates the development of data and knowledgedriven advanced techniques to extract insights and support informed decision-making. Within this context, the detection of anomalies, data points that deviate significantly from expected patterns, becomes crucial as these anomalies can arise due to diverse factors, including sensor errors, data corruption, and changes in underlying processes, all of which may impact system performance, accuracy, and overall efficiency. This paper thoroughly examines the diverse frameworks and architectures established for anomaly detection across multiple domains. It highlights the complexity associated with the nature of anomalies, which are often domain-specific and contextually bound, thereby presenting significant challenges in devising a universal framework capable of addressing anomalies regardless of the domain or context of the application. To address these challenges, the author proposes a comprehensive anomaly detection framework feature set (ADF2S) that captures the functional, structural, and operational dimensions of anomaly detection frameworks. The proposed anomaly detection framework feature set (ADF2S) and its crossframework evaluation contribute a practical foundation for researchers and practitioners, supporting the development of anomaly detection frameworks capable of balancing scalability, interpretability, and resilience.
This paper presents a method for improving the quality of multimodal "Text-Image" representations by integrating external ontological knowledge into a text query. The proposed approach integrates semantic and associative knowledge from ontologies directly into the text module of the CLIP neural network model through a tree-like syntax structure, similar to the K-BERT method. Adapting the syntax tree for use as an input parameter of the text module involves modifying the original positional encoding mechanism and attention mechanism. Relevant research was conducted on ways to modify the attention mechanism in accordance with the ontological nature of the syntax tree. The results of the experiments showed that the modified model outperforms the basic CLIP in terms of image classification accuracy in a number of object categories. A particularly noticeable increase in quality is observed for specific classes that require additional conceptual knowledge. The article discusses in detail the model architecture, the process of knowledge enrichment, and the impact of the proposed modifications on the model's performance.
Virtual reality (VR) has become a key tool in education, although some students still struggle with focus, leading to high dropout rates. The goal of this research was to create a more engaging and effective learning environment through immersive technology, but high costs remain a barrier for low-budget educational institutions. This study evaluates the implementation of an open-source immersive VR (IVR) solution for education, combining OpenSimulator, Firestorm VR, SteamVR and Meta Quest 2, as a cost-effective alternative to commercial platforms. Through systematic hardware benchmarking, we demonstrate a cost reduction compared to commercial solutions while maintaining high performance through optimized hardware configurations. Comparative analysis reveals that open-source solutions offer superior customization and privacy control, although they require greater technical experience for implementation. These findings provide actionable guidelines for institutions adopting IVR, balancing performance and affordability.
