Computer Science and Mathematics

Abstract: Model fusion integrates the capabilities from source models into a single target model. As the open-source AI ecosystem matures, Hugging Face has hosted more than 2M models. This growing pool provides a rich base for model reuse and capability integration. Yet existing surveys often cover only separate parts of this space, and they do not provide a unified definition or a systematic taxonomy. This survey defines model fusion and organizes prior work into three levels: parameter-level, representation-level, and behavior-level fusion. We also review related metrics, benchmarks, and applications, summarize current challenges, and identify future directions. Our goal is to provide a clear map of this area and support future work on model fusion. A comprehensive list of papers about model fusion is available at https://github.com/Baicaihaochi/Awesome-Model-Fusion-Survey.

Abstract: Large language models (LLMs) have revolutionized and have had significant impact on diverse domains such as healthcare, software development and autonomous systems by enabling natural language understanding and reasoning. However, single agent architectures limit their potential due to their non-collaborative nature and the reduced capability to perform complex, multi-disciplinary tasks which require teamwork, role division and adaptive decision making. To counter these shortcomings, Multi-Agent Systems (MAS) have been developed into a platform of contemporary artificial intelligence that allow the autonomous agents to interact, reason and communicate with dynamic and complex environments. Accompanied by the growth of experimentation and implementation of LLMs and Generative AI, MAS frameworks have been applied to real-world applications increasingly. This survey offers an in-depth and comparative study of three innovative and advanced MAS frameworks; Autogen, Langroid and MetaGPT. It delves into their architectural design, communication standards, scalability, applicability and their integration into the rising real world technologies. It presents standard benchmark criteria and performance measures (latency, throughput and memory utilization) through detailed case studies across diverse application domains such as e-commerce, medicine and AI-assisted software engineering. Moreover, it highlights important issues like explainability, security, computational cost and human-in-the-loop requirement in designing such models. Being a synthesis of theoretical developments and practical implementation experiences, it provides a systematic decision-making guide and serves as a basis of further MAS research and development.

Abstract: Selecting appropriate healthcare waste (HW) treatment technologies is a challenging multi-criteria decision-making problem characterized by uncertainty, conflicting evaluation criteria, and limited decision-support information. Existing approaches often rely on subjective weighting schemes and may provide rankings that are sensitive to variations in expert judgments. This study proposes ARAS-H-IW, a hybrid decision-support framework that combines Additive Ratio Assessment under Hesitant Fuzzy Sets (ARAS-H) with an Inverse Weighting (IW) mechanism capable of inferring criterion weights directly from expert preference constraints through constrained quadratic optimization. To evaluate its practical applicability, the framework was applied to a real-world healthcare waste management case study using data provided by the Regional Health Directorate of Fez-Meknes (Morocco). A fully reproducible Python-based web platform was developed to automate the complete analytical workflow, including hesitant fuzzy modeling, multi-expert ranking aggregation, inverse weight inference, comparative evaluation, sensitivity analysis, Monte Carlo robustness assessment, and automated reporting. The proposed framework identified centralized autoclaving as the most favorable treatment alternative, followed by regional outsourcing and microwave disinfection. Comparative analyses with TOPSIS, VIKOR, PROMETHEE II, and EDAS showed strong agreement regarding the best and worst-ranked alternatives. Sensitivity and Monte Carlo analyses further demonstrated the stability and robustness of the obtained rankings, while all expert aggregation strategies converged toward the same consensus ordering. The results highlight the capacity of ARAS-H-IW to generate transparent, reproducible, and robust decision recommendations under uncertainty. The proposed framework provides a practical tool for healthcare waste technology assessment and offers a promising foundation for supporting evidence-based decision-making in regional healthcare waste management.

Abstract: Evolutionary algorithms (EAs) are widely used nature-inspired optimization methods capable of solving complex and high-dimensional problems across science and engineering. Foundational paradigms such as genetic algorithms, genetic programming, differential evolution, evolution strategies, and evolutionary programming have expanded into multi-objective, surrogate-assisted, hybrid, and large-scale variants, broadening their applicability to dynamic and datadriven environments. This survey provides a structured review of EAs from a domain-centric perspective, focusing on how different techniques are designed for engineering problems. Applications are examined across renewable energy, civil and structural engineering, electronics, industrial optimization, healthcare, robotics, and smart cities. We present an updated taxonomy of classical and emerging algorithms, consolidate recent application studies, and review benchmarking and reproducibility practices essential for fair evaluation. Key challenges including scalability, constraint handling, and exploration–exploitation balance are discussed alongside future directions such as EA–deep learning integration, federated optimization, and interpretable evolution. This survey offers an updated view of EAs and their engineering relevance.

Abstract: Depression among university students has emerged as a significant mental health concern worldwide. Traditional assessment methods primarily rely on self-reported questionnaires and clinical evaluations, which may not provide scalable and continuous monitoring. Recent advances in machine learning have created opportunities to identify depression-related behavioral patterns through digital activity data. However, many predictive models operate as black-box systems that provide limited interpretability. This study presents a comparative analysis of explainable artificial intelligence (XAI) approaches for depression risk assessment using digital behavior data collected from university students. Logistic Regression and Random Forest classifiers were developed using behavioral indicators including screen time duration, social media usage frequency, nighttime device usage, sleep patterns, self-perceived digital dependency, and perceived academic impact. Model performance was evaluated using accuracy, precision, recall, and F1-score metrics. SHapley Additive exPlanations (SHAP) were applied to improve transparency and interpretability. Experimental results indicate that Logistic Regression achieved 94.74% accuracy, while Random Forest achieved 100% accuracy on the testing dataset. SHAP analysis identified academic impact as the most influential predictor of depression risk. The findings demonstrate that explainable machine learning models can support transparent and ethical depression risk assessment in higher education environments.

Abstract: Two independent runs of the full metric suite yield: (a) full plaintext avalanche from round 1 (mean 63.97–64.67 of 128 bits, ideal 64); (b) the differential-probability sampling floor of 2×10⁻⁵ reached at round 4 (50,000 of 50,000 output differences distinct, both sessions); (c) algebraic-degree lower-bound saturation at the maximum observable value from round 1; (d) linear bias indistinguishable from random (combined exceedance 4.40%, below the 4.55% noise floor); and (e) branch numbers at the Singleton (MDS) bound for every tier (B = 5 for 4×4, B = 9 for 8×8, B = 17 for 16×16), computed exhaustively over weight-1 inputs. MD-Hill-SPN therefore moves beyond theoretical construction to a construction that passes a defined empirical evaluation suite avalanche, differential sampling, linear-bias probing, algebraic-degree lower bounds, and MDS branch numbers under single-key, known-plaintext conditions with fixed parameters, an evaluation no prior Hill cipher variant has reported in full.

Abstract: The advent of Cloud-Fog-Edge computing has transformed distributed data processing by performing computation closer to end devices. Due to resource constraints at edge nodes and the dynamic nature of fog-assisted communication, secure and efficient group key distribution and batch verification in such decentralized systems remain a major challenge. Many existing protocols based on Chinese remainder theorem (CRT) use straightforward scalar product with no non-linear transformation to mask the group key and hence fail to provide semantic security. Others suffer from architectural overhead since they require distinct and independent sets of moduli equations with multiple mathematical structures for different network layers, which increases computing overhead, limits scalability and synchronization delays during frequent node leave/join. To mitigate these challenges, this paper proposes a unified distributed CRT-based protocol for Cloud-Fog-Edge environments. Our protocol introduces a non-linear two-factor encryption logic, by incorporating a unique secret parameter for every edge node and breaks the algebraic linearity of the CRT. Additionally, our protocol uses a single set of moduli equations across Cloud-Fog-Edge networks which drastically reduces computation and storage costs at the fog layer. Our protocol achieves O(1) efficiency for rekeying. Formal security analysis using ProVerif and ROR model demonstrates that our protocol has considerable security advantages. To prove its practicality, an ESP32-based simulation on Wokwi is used to verifies the correctness of group key distribution, retrieval, and batch message verification. The performance analysis findings show that our protocol outperforms others in computation cost, communication cost, security and applicability for resource-constrained Cloud-Fog-Edge computing networks.

Abstract: Tennis match prediction has been studied extensively, yet the literature offers no controlled comparison of Elo ratings, classical machine learning, and deep neural networks under identical experimental conditions, leaving practitioners without clear guidance on model selection. We address this gap with a unified empirical study on 133,138 professional men’s tennis matches from the Association of Tennis Professionals tour (1968–2024). Four approaches are evaluated on the same temporally split data with a common 16-feature set and an aligned evaluation protocol: an enhanced Elo rating system, ten classical machine learning algorithms, seventeen deep neural network configurations spanning 207,000 to 21,000,000 parameters, and a hybrid Elo–machine learning (ELO-ML) approach that augments classical learners with three Elo-derived features. A tuned Elo baseline alone reaches 65.87% accuracy, the best of ten classical machine learning algorithms reaches 66.30%, seventeen deep neural network configurations cluster at 66.15–66.22%, and the hybrid ELO-ML approach reaches 67.52% (McNemar’s test, p < 0.001 for all ELO-ML pairwise comparisons). All four approaches sit within a 1.65 pp band whose upper edge lies below the 70–72% accuracy commonly cited for bookmaker odds, indicating that pre-match prediction under universally available features is a difficult task in which Elo alone already captures most of the predictable signal and algorithmic sophistication adds only marginal headroom. Deep neural networks deliver substantially better probability calibration than the other approaches (Expected Calibration Error 0.0077 vs. 0.0142). Model capacity exhibits sharply diminishing returns: all seventeen network configurations, spanning a 100-fold range in parameter count (207,000 to 21,000,000), fall within a 0.07 pp accuracy band. The study establishes a controlled benchmark for tour-level tennis prediction, quantifies how narrow the headroom above Elo actually is, provides modest but consistent empirical support for the Statistically Enhanced Learning framework, and supplies deployment-ready operating points for sports analytics practitioners.

Abstract: Human activity recognition (HAR) based on smartphone and wearable sensor data is commonly addressed using statistical learning methods and deep neural networks that often provide strong predictive performance, but at the expense of limited interpretability and substantial computational and energy requirements. Such limitations reduce their suitability for deployment in practical sensing environments where model decisions must be transparent, verifiable and executable on resource-constrained devices. In this work, we investigate the Convolutional Tsetlin Machine (CTM) for multimodal HAR using the UCI-HAR dataset. The Tsetlin Machine is a novel neuro-symbolic machine learning approach that offers two important advantages over many conventional machine learning methods: (i) it learns logic-based decision rules that are human-readable and formally verifiable, and (ii) it operates with comparatively low computational complexity, making it well suited to efficient and low-power on-device learning. The proposed study systematically analyses the contribution of different feature modalities by decomposing the inertial signals space into semantically defined subsets according to: (i) sensor source: accelerometer or gyroscope; (ii) physical component: body or gravity; (iii) coordinate: x, y or z. A separate CTM classifier was trained for each modality and their combination in order to determine the relative discriminative value of each modality group for activity classification. In addition to predictive performance the study emphasizes the interpretability of the CTM model ensured by expressing each decision in the form of propositional clauses, thereby enabling visualization and direct inspection of the modality-specific patterns supporting each activity class. Owing to its symbolic structure and modest computational demands, the CTM provides a principled framework for the design of explainable, resource-efficient and deployable HAR systems. The proposed work therefore contributes toward trustworthy multimodal sensing by jointly addressing predictive performance, interpretability and suitability for embedded and mobile platforms.

Abstract: Air quality forecasting and environmental health research at urban and regional scales depend on the combination of measurements from heterogeneous sensor networks, yet the construction of integrated multi-source datasets is rarely described or released as a self-contained deliverable. This paper presents an open dataset that combines four sensor-derived sources covering the whole of Spain over the period 2022 to 2024: hourly air quality observations from the 588 stations of the national network operated by the Ministerio para la Transición Ecológica y el Reto Demográfico (MITECO), daily meteorological records from the Agencia Estatal de Meteorología (AEMET), daily mobility indicators derived from anonymised mobile telephony events published by the Ministerio de Transportes y Movilidad Sostenible (MITMA) at the municipality level, and a calendar of national and Autonomous-Community public holidays. The processing pipeline harmonises sources that differ in temporal resolution, spatial codification and quality regime into a tidy hourly table indexed by station and timestamp, with a fixed feature schema of 56 variables per record. Air quality stations are paired with their nearest AEMET station through a three-tier distance rule, and the daily exogenous features are aligned to the air quality time axis through a two-variant temporal-alignment scheme (lag-and-expand to the hourly grid for the hourly release, same-calendar-day join for the daily release). A complementary daily-resolution variant of the dataset is also released, with 72 columns and the same feature schema except for the air quality block, which is aggregated to daily mean, minimum and maximum. The integrated dataset contains approximately 14 million hourly records across the 588 stations and is released on Zenodo (DOI 10.5281/zenodo.20196221) under a Creative Commons Attribution 4.0 International (CC BY 4.0) licence. It is intended as a substrate for research on air quality forecasting, environmental epidemiology and multi-source data fusion at nationwide scale.

Abstract: Aspect-based sentiment analysis (ABSA) is increasingly the granularity at which customer feedback is consumed, and recent work has pushed the field rapidly toward transformer- and graph-based architectures [3,5–7]. However, most modern ABSA approaches assume either a closed manually curated aspect taxonomy or a fully supervised aspect extractor trained on benchmark corpora such as SemEval. Neither assumption holds in low-resource emerging-market settings, where aspects must be discovered from the corpus itself, annotation budgets are negligible, and class distributions can be unexpectedly skewed. This article introduces Soft-Aspect ABSA, a probabilistic, topic-model-agnostic framework that promotes unsupervised topic-model output to first-class aspects via a temperature-controlled softmax over topic-membership posteriors. We instantiate the framework with a spectral-clustering plus non-negative matrix factorisation (NMF) substrate on a corpus of 292 Google Play Store reviews of a Ghanaian retail-bank mobile application (April–September 2024). The corpus exhibits an inverted class imbalance (30.8% positive / 69.2% negative under a keyword-bootstrap rule) and a four-cluster topic decomposition. A baseline TF-IDF embedding head trained with binary cross-entropy collapses to the majority class on the held-out test set: accuracy 0.6949, minority-class F1 0.000, Matthews correlation 0.000, despite a ROC-AUC of 0.934 that indicates well-ranked probabilities. The framework licenses two closed-form remediations — class-weighted cross-entropy and focal loss [28] — that we evaluate empirically on the same head. Focal loss with γ = 1 lifts minority-class F1 from 0.000 to 0.818, Matthews correlation from 0.000 to 0.746, and ROC-AUC to 0.986, demonstrating that the framework correction is not merely formal but is recoverable on the case-study data. We also run a bootstrap stability protocol for cluster-count selection (B = 50) that flags the silhouette-max k* = 4 as only moderately stable (I_stab = 0.64). The contribution is methodological: a reusable scaffold for low-resource ABSA pipelines in which the aspect set is not given a priori.

Abstract: Phishing is the most common cybersecurity threat. With phishing, attackers create a website or manipulate a URL in order to obtain a user’s sensitive information. Sensitive information can include a user’s credentials, payment details, or personal information. Phishing attacks target online users by baiting them to click on a fraudulent link. Phishing is a growing concern for users across the world. I propose a phishing detection framework that is lightweight, fast, and able to detect URLs with phishing content. The lightweight comparative phishing framework focuses on the extraction of a reduced number of URL features. These features include characteristics, structures, and patterns that are seen in URLs. These features prepare and place input to the three supervised machine learning methods: Logistic Regression, Decision Tree, and Random Forest. The frameworks were then evaluated based on four main classification metrics: accuracy, precision, recall, and F1-score. The Random Forest Classifier, within the lightweight comparative machine learning framework, was the most accurate in phishing detection with minimal computational requirements. The purpose of this lightweight framework was to offer real time cyber security solutions on browsers. The framework was scalable and efficient.

Abstract: This paper introduces a Convolutional Neural Network (CNN) to jointly classify images with multiple classes on the Fashion-MNIST dataset, with a test accuracy of 90.20% and 0.11 million parameters of parameters a lightweight model, which significantly outperforms classical baselines (HOG+SVM: 85%) and is both computationally efficient. The CNN uses three convolutional blocks with varying filter depth (3264128), ReLU activation, MaxPooling, Batch Normalization, Dropout regularization, and fully connected classification head that is trained using Adam optimizer. These architectural concepts are generalised to the field of AI-related cybersecurity: namely, the deep learning-based Network Intrusion Detection Systems (NIDS) classifying network traffic flows as benign and attack ones - a problem that is characterised by the same core challenge architecture as Fashion-MNIST (spatial feature hierarchy extraction, multi-class discrimination, imbalanced class difficulty). State-of-the-art CNN based IDS are 94.8-97.5% accurate in detection (Attention-CNN-LSTM; Nature Scientific Reports, 2025), 98.5% with combined host/network data (Springer Nature, 2024), and 99.67% with encrypted malicious traffic.

Abstract: The concept of Z-number was introduced to formalize partially reliable information. A Z-number represents linguistic evaluations of a random variable's value and the associated reliability degree. The latter is defined as a fuzzy restriction on the value of a probability measure since the actual probability distribution is unknown. Lotfi Zadeh formalized an extension principle for computation with Z-numbers based on fuzzy and probabilistic arithmetic and noted that the problem of computing with Z-numbers is easy to formulate but difficult to solve. Since then, a series of theoretical studies and practical applications of Z-numbers has been proposed. However, the computational complexity of Z-numbers remains a challenge. Because the actual probability distribution is unknown, a set of probability distributions is considered, which is the main source of computational complexity. In this study, we outline a new approach to computation with Z-numbers that relies on the concept of imprecise probability. Specifically, we use a lower prevision measure (the lower envelope of a set of probability measures) as the basis for computation. The reason is a one-to-one correspondence between lower previsions and convex sets of probability measures. Experimental results show that the proposed approach reduces computational complexity compared with existing methods.

Abstract: Electric vehicle routing problems (EVRPs) have become increasingly important in sustainable logistics, where routing decisions must account for limited battery capacity, charging constraints, and time windows. In such settings, fast and adaptive decision-making is critical. Nature-inspired genetic programming (GP)-based hyper-heuristics, particularly those based on routing policies, offer a promising approach by evolving decision rules through mechanisms analogous to biological evolution, enabling adaptive behavior in dynamic environments. However, existing GP-based approaches have been developed for deterministic routing settings, limiting their applicability in real-world environments characterized by uncertainty. This work extends a GP-based route generation scheme (RGS) framework to a stochastic electric vehicle routing problem with time windows (EVRPTW) setting by incorporating variability in customer demand, service time, and travel time through controlled stochastic perturbations. The priority function representation is extended with additional terminals capturing stochastic effects and global system information, and new variants of existing routing strategies are introduced to improve robustness through capacity-aware vehicle selection. The proposed method is evaluated across a range of stochastic scenarios and objectives, including fleet size, energy consumption, and total tardiness. The results demonstrate that different routing strategies are best suited to different optimization objectives, with each approach consistently exhibiting strengths aligned with specific performance criteria, while stochasticity primarily amplifies these structural differences rather than altering their relative behavior. These findings further support the suitability of evolutionary, nature-inspired approaches for constructing adaptive decision policies in stochastic routing environments.

Abstract: Visual programming using blocks and diagrams facilitates understanding of fundamental programming concepts, which is particularly important for children with special educational needs because it reduces their cognitive load and encourages interactive learning. This study aimed to develop and apply a hybrid multi-criteria framework to evaluate, rank, and select visual programming software solutions intended for children with special educational needs. Based on an analysis of the educational context and the target population’s needs, a set of criteria was defined to evaluate and select the most suitable software solution. Data for the analysis were collected using a structured questionnaire, from which a decision matrix was developed. Within the proposed hybrid multi-criteria decision-making (MCDM) framework, criterion weights were determined using the Rank Order Centroid (ROC) method, and the ranking of alternatives was performed using the Preference Ranking Organization Method for Enrichment Evaluation (PROMETHEE II). Additionally, a sensitivity analysis was conducted to assess the stability and robustness of the obtained rankings in relation to changes in the criterion weights. The results indicate a stable ranking of alternatives and the identification of the most favorable solution in the majority of scenarios. The projection quality of 91.1% in the Geometrical Analysis for Interactive Aid (GAIA) plane confirmed the reliability of the visual interpretation of the results. The proposed framework improves the decision-making process and provides a foundation for further research in educational software evaluation.

Abstract: Brain-Computer Interface (BCI) is a system that enables the human brain to interact with the external environment without using muscles. Electroencephalography (EEG) is a commonly used modality for measuring neural brain activity. However, its low signal-to-noise ratio (SNR) and electrode reference problems lead to poor spatial resolution. As a result, EEG signals are often contaminated with physiological artifacts such as muscle movements. Therefore, this study used novel tripolar concentric ring electrodes (TCREs) that are used to record brain activity related to the overt and covert speech. Brain signals associated with overt and covert speech were recorded using TCREs and disc electrodes. Classification algorithms, including K-Nearest Neighbors (KNN), Fully Connected Neural Networks (FCNNN), and Convolutional Neural Networks (CNN), were used to classify the TCRE and conventional EEG signals. The data was collected from 16 healthy participants. The experimental results demonstrate that TCREs provide superior performance compared to conventional disc electrodes. In addition to that, the 0.5–1.2s interval, corresponding to the peak stimulus window, exhibits a maximum power of 250µV. The average accuracy achieved during this epoch was 86.25%, whereas the remaining epoch shows an accuracy of 83.5% using TCREs.

Abstract: The rapid advancement of quantum computing poses a fundamental threat to classical public-key cryptographic systems, necessitating the transition to post-quantum cryp-tography (PQC). While significant progress has been made in the standardization of quantum-resistant algorithms, their practical deployment in heterogeneous environ-ments—particularly resource-constrained Internet of Things (IoT) devices—remains a critical challenge. This study presents a comprehensive experimental evaluation of four NIST-standardized PQC algorithms: CRYSTALS-Kyber (ML-KEM), CRYS-TALS-Dilithium (ML-DSA), FALCON, and SPHINCS+. The analysis is conducted across two distinctly different hardware platforms: a high-performance x86-64 system (AMD Ryzen 7 5700U) and a resource-constrained embedded microcontroller (ESP32-WROOM). Implementations were developed in C, Go, and Python to assess the influence of programming environments on algorithmic efficiency. The evaluation focuses on key performance indicators, including computational la-tency, memory consumption, communication overhead, and temporal determinism, based on extensive benchmarking over 1,000 iterations. Experimental results demon-strate significant trade-offs between security level, execution performance, and re-source utilization. Lattice-based schemes such as Kyber and Falcon exhibit superior ef-ficiency and scalability on embedded platforms, while Dilithium encounters stability limitations due to memory constraints. In contrast, SPHINCS+ provides strong security guarantees and architectural robustness at the cost of higher computational overhead. The findings highlight the critical role of hardware-specific optimizations and soft-ware design choices in enabling practical PQC deployment. This work provides ac-tionable insights for selecting and tuning post-quantum algorithms in real-world sce-narios, supporting the secure migration of IoT and distributed systems toward quan-tum-resilient infrastructures.

Abstract: Construction companies, petrochemical companies, and airports are classified as large-scale organizational and technical systems. In organizational and technical system is implemented using hierarchical distributed control systems. To achieve goals, numerous parallel technological operations and business processes are executed, requiring synchronization and consuming significant energy resources. To optimize energy resources at all hierarchical levels, an up-to-date picture of the system state is necessary. This paper proposes constructing a snapshot of the system's state, which allows for an assessment of system performance within selected criteria and the selection of appropriate solutions within the entire control system. Procedures for constructing a snapshot are discussed. An example of optimizing the energy consumption of a control system based on snapshot analysis is provided.

Abstract: Foundation-model agents now use reusable skills for tool use, long-horizon planning, and adaptation across related tasks. The term, however, is used loosely. It may describe a prompt package, an executable workflow, a learned routine, or an artifact distributed through a repository. That looseness makes it hard to compare methods, measure progress, or discuss security and governance with precision.We study agent skills as reusable and adaptive units of competence between model capability and situated task execution. The survey separates skills from nearby constructs such as prompts, tools, memory, and policies, then organizes the literature around representation, lifecycle and orchestration, evaluation, security and governance, and application domains. Across these areas, skill quality is only one part of the story. Useful skills also depend on abstraction choices, retrieval and composition mechanisms, ecosystem structure, and infrastructure security. We treat agent skills as a research object in their own right and identify open problems in automatic induction, cross-environment transfer, longitudinal evaluation, and trustworthy sharing in open agent ecosystems. A public paper list is available at https://github.com/JinhaoShen/awesome-agent-skill-papers.