Today, simply reducing carbon emissions is not enough to combat climate change; the existing carbon dioxide stock in the atmosphere must also be effectively eliminated. However, the factors that determine the success of investments in this area and the most appropriate financial models to support these investments have not been systematically analyzed in the literature. This deficiency causes decision-makers to experience uncertainty in resource allocation and reduces the effectiveness of strategic planning. To address this gap, this study develops a model that integrates parameter-driven synthetic evaluation to expand expert opinions, consensus-based expert selection with Manhattan distance-based centrality to objectively determine expert weights, and cognitive maps to analyze criteria relationships. Moreover, a dynamic multi-facet fuzzy sets approach is developed to dynamically model uncertainties, introducing a new fuzzy set structure to the literature. Thanks to these methodological innovations, the model outperforms existing models in terms of both technical accuracy and analytical depth. According to the research findings, “monitoring quality” and “policy incentives” are the most critical criteria for investments in net-negative emissions technologies. Among the investment alternatives, “insurance funds for early-stage net-negative emissions technologies” and “bonds tied to verified carbon removal performance” are identified as the most effective innovative investment models. In conclusion, this research fills a significant gap in the literature, both theoretically and practically, and provides an innovative methodological contribution to decision-support systems for net-negative emissions technologies within the framework of sustainable finance.
The Markov decision process (MDP) provides a framework for sequential decision optimization under uncertainty and serves as a foundation for dynamic programming problems. However, current robust optimization approaches often overlook the social attributes of decision-makers (DMs), while behavioral models that do account for social factors often treat them as external constraints rather than integrating them into a distributionally robust optimization (DRO) framework. To address these challenges, we construct a distributionally robust Markov decision process (DRMDP) that integrates empathetic relationships, aiming to minimize the system's total cost while also improving decision-making stability. First, we model DM profiles and the dynamic evolution of a group trust network to capture the effects of subjective performance evaluation. Next, we establish a principled learning process to generate a global empathetic weight matrix. Then, we construct a psychological cost mechanism that quantifies deviations from group consensus and trades off economic objectives against social alignment. Subsequently, we design an improved DRO process to analyze the stability of the MDP under empathetic relationships. Finally, experiments reveal a synergistic effect: the empathetic mechanism mitigates the conservatism of pure robust strategies, leading to a substantial reduction in both average total cost and cost dispersion. These findings demonstrate that high-quality group consensus can complement the data-derived robustness enhancing both economic efficiency and decision stability.
Slacks-based measure (SBM) is a non-radial DEA model that considers the assumption of non-proportional changes in resources and products. The traditional DEA models cannot differentiate between efficient units. Although the super-efficiency model has been widely used in literature to distinguish among efficient units, its use for ranking has been criticized by some scholars. This paper proposes an integrated model to simultaneously evaluate both efficient and inefficient DMUs within a unified framework. Although previous studies have proposed SBM-based models that attempt to rank all DMUs simultaneously, some of these approaches may suffer from infeasibility or lack of discriminatory power. This paper proposes a novel model to enhance ranking discrimination while ensuring feasibility and interpretability. A theoretical branch analysis shows that the proposed formulation can be implemented by solving n + q linear programs for all n DMUs, where q ≤ n is the number of evaluations requiring the super-efficiency branch. This LP count and the reported wall-clock times refer to the Phase-1 integrated evaluation only and do not include the subsequent Phase-2 projection adjustment. Controlled experiments covering six data configurations and 20 independent replications show that the proposed Phase-1 implementation attains the lowest median total wall-clock time in all tested configurations. Relative to the sequential SBM-SupSBM procedure and the integrated models of Tone et al. and Lee, the reductions in median computational time range from (0.9%) to (10.5%), from (30.0%) to (75.6%), and from (25.1%) to (71.3%), respectively. These empirical findings are restricted to the reported implementation and computational environment. Guaranteed feasibility, unit invariance, and independence from a user-selected big-(M) parameter are the main structural properties of the proposed model. In addition to integrating the SBM and SupSBM models, the extended framework applies a second-stage Pareto adjustment to weakly efficient initial projections. Its additional value lies in combining this adjustment with the proposed parameter-independent first-stage model and in measuring the inefficiency of the initial projection through an SBM-type normalized fractional term incorporated into the final evaluation.
Photovoltaic (PV) modules are increasingly deployed in large-scale renewable energy systems, making reliable defect detection and fault diagnosis essential for operational safety, energy yield, and maintenance efficiency. This review systematically examines computer vision-based approaches for PV module inspection from a decision-oriented perspective. First, it summarizes representative PV defects, including cracks, hot spots, soiling, shading, delamination, corrosion, broken cells, finger interruptions, burn marks, and bypass diode-related anomalies, and discusses their visual manifestations, physical origins, diagnostic implications, and maintenance relevance. Second, it reviews major imaging modalities, including electroluminescence imaging, infrared thermography, RGB imaging, and UAV-based inspection, highlighting their complementary roles in detecting internal, thermal, and surface-level defects. Third, it analyzes key computer vision tasks, such as image classification, object detection, semantic and instance segmentation, anomaly detection, and severity assessment. The review emphasizes that PV inspection should move beyond normal/defective recognition toward quantitative and trustworthy fault diagnosis. Finally, it discusses current challenges and future perspectives, including field robustness, data scarcity, small-defect detection, multimodal fusion, explainability, uncertainty estimation, edge deployment, and human-in-the-loop maintenance decision-making. This review provides a structured reference for developing intelligent and actionable PV inspection systems.
Group decision-making is ubiquitous in human society, aiming to pool individual wisdom to address complex problems. Traditional group decision-making prioritizes achieving group consensus, that is seeking a compromise solution acceptable to all decision-makers. However, in predictive and judgmental tasks with objective ground truths, such as expert forecasting and risk assessment, the goal of group decision-making should shift from orchestrating agreement to identifying the single optimal decision that best corresponds to reality. This paper proposes a non-consensus aggregation algorithm that minimizes the correlation between decision-maker weights and decision deviations, thereby guiding aggregated results toward the optimal decision. Theoretically, the method is proved to provide the maximum likelihood estimate of the optimal decision. Extensive experiments on multi-problem decision-maker judgment datasets demonstrate that the proposed method significantly reduces decision errors compared to conventional aggregation methods, particularly for difficult or large-scale problems. This work offers a new and effective path for group decision-making in accuracy-oriented scenarios.
Combining multi-criteria decision-making (MCDM) with Geographic Information Systems (GISs) provides a substantial framework for spatial suitability analysis, which assesses locations based on environmental, economic, and social criteria. The aim of the study is to highlight the research advances in this field, and to detect the most usable methods in each subject area. We followed a semi-systematic review approach, combining bibliometric analysis of 486 articles with a manual content analysis of a stratified sample of 150 studies. The bibliometric analysis was performed on three major terms: “MCDM”, “Suitability”, and “GIS”, from 2000 to 2024. The obtained results reveal that The Most used method is the analytic hierarchy process (AHP). Finally, the study provides field-specific methodological frameworks highlighting the most commonly adopted approaches in the literature to serve as a reference for future studies.
This paper presents an advanced autonomous cyber-defense framework for IoT-edge networks, leveraging a novel integration of Starfish Algorithm-Optimized Bayesian Constitutive Neural Networks (BCNN) within Cross-Domain Zero Trust architecture. The system begins with input data sourced from two comprehensive and diverse datasets, IoT-24 and TON_IoT-v2, which reflect a wide range of IoT traffic patterns and attack scenarios. To enhance data quality, a Reverse Lognormal Kalman Filter (RLKF) is employed for pre-processing, effectively normalizing and denoising the input data. The cleaned data is then analyzed by BCNNs, which provide probabilistic modeling to account for uncertainties in cyber threat detection. To overcome the inherent limitations of BCNNs in parameter tuning, the Starfish Optimization Algorithm (SFOA) is used to optimize the model, significantly improving detection accuracy while minimizing computational demands—an essential feature for deployment in resource-constrained edge environments. Security is further reinforced through a Cross-Domain Zero Trust model, enabling continuous authentication and verification of data flows to prevent unauthorized access. The proposed technique attains 8.26%, 3.06%, and 3.00% higher accuracy, 3.48%, 3.13%, and 3.39% higher precision over state-of-the-art methods including QIFL, ZTA-IoT, and XAI-based detection systems. The proposed approach thus delivers a scalable, intelligent, and resilient solution for securing heterogeneous IoT-edge ecosystems.
In recent years, non-coherent multiple-input multiple-output (MIMO) detection in rapid fading channels have gained much interest but is less affected by the issues like phase variations and has minimal synchronization requirements and channel estimation. In MIMO systems, the users exchange information via relay nodes that use maximum ratio combining/maximum ratio transmission. Hence, the users are unaware of the direct link, and thus it is treated as interference. This work adopts an optimized block-diagonal coded M-ary frequency shift keying (BD coded MFSK) modulation with an energy detection approach for handling direct-link interference. The proposed Energy detector with Inference cancellation (EIC) specifically designed for MIMO-MFSK (i.e. EIC-MIMO-MFSK) does not require any model for estimating the channel state information (CSI), and thereby the overhead of the system is decreased. Also, the conventional MFSK modulation does not monotonically enhance the Symbol error rate (SER) performance while increasing the modulation level. To tackle this issue, a new Aquila optimizer is introduced that minimizes the SER by detecting the optimal value of modulation levels. Furthermore, a novel block-diagonal (BD) pre-coding is developed before modulation to eliminate interference from other users. Finally, the simulation results are carried out by comparing them with existing systems. The proposed system reduces the SER performance as compared to others. The simulation results indicate that the proposed EIC-MIMO-MFSK attains a minimum SER between 10-15 dB SNR and is the best result compared to state-of-the-art techniques.
Cloud Computing (CC) becomes a fundamental platform for implementing large-scale and dynamic workflows; nevertheless, efficient resource utilization remains a critical challenge due to fluctuating workloads, energy constraints and heterogeneous virtual machine (VM) environments. To address these issues, an Optimized Gated Fusion Adaptive Graph Neural Network dependent Scheduling Framework in Cloud Computing for Efficient Resource Utilization (GFAGNN-SFER-CC-BKOA) is proposed. The framework integrates security, intelligent task clustering, adaptive resource monitoring, and meta-heuristic optimization to improve workflow scheduling performance. Initially, user authentication is ensured through a UUID-BLAKE based hashing mechanism, enabling secure access to cloud resources. Workflow tasks are then clustered using Localized Sparse Incomplete Multi-view Clustering (LSIMC) to reduce makespan and scheduling overhead. A Gated Fusion Adaptive Graph Neural Network (GFAGNN) is employed to monitor and predict VM resource availability by capturing dynamic spatio-temporal dependencies in cloud workloads. Finally, the Black-Winged Kite Optimization Algorithm (BKOA) selects optimal VMs for dynamic workflow scheduling, aiming to maximize resource utilization while decreasing energy consumption, cost and execution time. The proposed framework is evaluated using the GWA-T-12 Bitbrains dataset and implemented in Python. Experimental results demonstrate that GFAGNN-SFER-CC-BKOA significantly outperforms existing methods, including MOSF-A-RNN-CC, PMRP-DCRNN-CC and MOPWS-DQN-CC, achieving high accuracy of 99.7%. These outcomes confirm the efficacy of the GFAGNN-SFER-CC-BKOA in cloud computing environments for secure, adaptive and energy-efficient workflow scheduling.
Ultrasound imaging has become an important tool for breast cancer screening. With the growing application of deep learning in medical imaging, numerous methods have been developed to enhance the accuracy and efficiency of breast ultrasound interpretation. However, a systematic understanding of these approaches remains limited. This review aims to comprehensively evaluate the development of deep learning techniques for breast ultrasound-based auxiliary diagnosis. Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, a structured search was conducted across 11 academic databases to identify studies published between January 2016 and October 2024. A total of 2,779 records were initially retrieved. After duplicate removal, eligibility screening, and quality assessment, 65 studies were included in the review. These studies were categorized into three main subtopics: 24 focused on lesion classification, 14 on object detection, and 27 on lesion segmentation. The categorization reflects the primary clinical tasks in breast ultrasound diagnosis, including identifying lesion types, localizing lesion regions, and delineating lesion boundaries. Recent evidence shows that deep learning-based methods have achieved promising performance through strategies such as feature enhancement, multi-scale feature extraction, and few-shot learning. Finally, this review discusses future work and challenges in applying deep learning to auxiliary diagnosis in breast ultrasound imaging. This paper presents a systematic review that summarizes current studies on deep learning-based auxiliary diagnosis in breast ultrasound imaging.
The multiple criteria decision-making model has gained significant attention in various fields. However, constructing a model that can focus on risk aversion behavior and incomplete information in multiple criteria decision-making is a challenging endeavor. Simultaneously, uncertainties such as human evaluation biases in multi-criteria decision-making are often unavoidable. To address these research gaps, we develop a novel risk-averse robust strategic manipulation approach with incomplete information under uncertain environments in this paper. First, we characterize the decision maker's risk aversion behavior based on the Mean-Variance theory and propose a strategic manipulation model with risk aversion behavior. Second, we utilize this strategic manipulation model as a benchmark and further integrate incomplete weight information. Furthermore, we employ a robust optimization method to capture evaluation biases in both the mean and covariance of compensation cost under uncertainty, ultimately proposing a novel risk-averse robust strategic manipulation approach with incomplete information. Through duality theory, the approach is equivalently transformed into a series of computationally efficient convex programming models. The proposed model ensures that holistic preference information can be acquired based on incomplete weight information constraints, thereby enhancing the flexibility of the decision-maker in setting preferences. Finally, the practicality of the proposed approach is demonstrated by applying it to the circular economy development problem, financial performance analysis and healthcare supplier selection. Comparative analysis and sensitivity analysis confirm the good performance of the proposed model in handling risk cost, uncertainty, and incomplete information.
To address the current lack of clarity in the distribution of targeted munitions, we propose a multi-target weapon target allocation model that considers shooting efficiency and construct a novel algorithmic framework. In this framework, we introduced the dominance matrix and developed our own adaptive genetic selection mechanism, and proposed a non-dominated sorting genetic algorithm (D-A-NSGA-III) based on the dominance degree matrix and adaptive genetic selection mechanism. The model clearly defines the correlation between the probability of damage to the target and the amount of munitions allocated, based on the shooting efficiency, and establishes a clear munitions allocation strategy for each target with a corresponding weapon-target allocation scheme. D-A-NSGA-III improves the convergence speed of the algorithm by reducing redundant comparisons in the ranking scheme, effectively preserving the structure of the good non-dominated solutions and improving the inferior dominated solutions. The results of the numerical experiments show that D-A-NSGA-III reduces its convergence time by a maximum of 77.63% and a minimum of 20.09% under three different experimental scenarios compared to the non-dominated degree-ranked genetic algorithm based on the reference point mechanism (NSGA-III) and the non-dominated sorting genetic algorithm based on dominance matrix (D-NSGA-III). This significant improvement not only validates the reasonableness of the proposed model, but also fully demonstrates the superiority of the D-A-NSGA-III algorithm over other similar algorithms. This research not only provides an effective technical approach to unmanned air defence artillery operations, but also offers a new solution to the multi-objective optimization problem.
Proportional fuzzy sets (PFSs) address real-world uncertainty using proportions instead of exact decimal values, offering a more flexible and intuitive approach to represent partial membership. This paper introduces proportional Fermatean Fuzzy Set (PFFS), an extension of Fermatean FS (FFS). PFFSs use proportional relationships to determine membership and nonmembership degrees (NMD), thereby improving accuracy and reliability under uncertain environments. This paper combines PFFSs with the TOPSIS method, a widely used MCDM technique. The PFFS-TOPSIS combination offers a more effective decision-making process, especially in cases where the criteria are vague or uncertain. To demonstrate its practical application, the paper uses the PFFS-TOPSIS method in a clinic selection problem, where criteria like cost, quality of service, and location may have imprecise values. Sensitivity and comparative analyses validate the PFFS-TOPSIS method, demonstrating its robustness, flexibility, and superior performance in handling uncertainty, accuracy, and efficiency compared to traditional decision-making techniques in complex problems.
Agriculture is a predominant occupation in India, and the country’s GDP is influenced by agricultural production. Agricultural failure can affect production in other sectors leading to a decrease in the country’s GDP. Nearly 70% of the population relies on agriculture for livelihood and employability. The area of arable land is diminishing due to the tremendous increase in population. Also, farmers lack the technical knowledge to overcome the challenges such as variable climatic conditions, unhealthy soil due to repeated cultivation in the same land, floods and other unexpected natural disasters. The success of agricultural development begins with identifying the suitability of land for cultivation, and it is essential for sustainable development. The suitability of arable land can be evaluated using several parameters such as soil and water. Thus, the Agriculture Land Evaluation Model (ALEM) is proposed in this paper to assist farmers in identifying the suitability of their agricultural land for crop cultivation by considering several factors. Various methods, such as Shannon’s Entropy Method, PROMETHEE and the soft set algorithm, have been applied for the development of the mathematical model. Shannon’s Entropy Method is used for calculating the priorities of the evaluation factors considered for decision-making. The weighted PROMETHEE method is applied for generating the preference values of the agricultural dataset, and the bijective soft set algorithm is used for the generation of classification rules. The developed model is validated using various farm datasets and opinions from experienced professionals in the relevant field. Furthermore, the model achieved values above 90% across the evaluated metrics when compared to other machine learning algorithms like Naïve Bayes, KNN, Decision Tree, Random Forest and SVM. Thus, ALEM produced better results for the given problem and will provide greater insights for researchers in the field.
A pandemic outbreak emerged in Wuhan, China, in December 2019. Shortly after that, the World Health Organization (WHO) identified the causative agent as a novel member of the Coronavirus family. Genetic analysis indicated that the SARS-CoV-2 virus was closely related to severe acute respiratory syndrome (SARS). The disease caused by SARS-CoV-2 is termed COVID-19. This study analyzes hemogram parameters and COVID-19 Polymerase Chain Reaction (PCR) test results using various machine-learning (ML) techniques. The ML techniques applied are Support Vector Machine, Adaptive Boosting, Gradient Boosting, Light Gradient Boosting Machine, Extreme Gradient Boosting, and Random Forest. In addition, particle swarm optimization and genetic algorithm are used for hyperparameter optimization. Both heuristic algorithms demonstrated capable search trajectories in optimizing ensemble tree parameters, effectively mapping the complex nonlinear optimization landscape. It was observed that using only the ten most valuable parameters did not significantly affect the results. Additionally, a trapezoidal membership function was created to simulate decision-making processes. The original dataset was compared with a dataset modified using trapezoidal membership functions, with degrees set according to the decision-makers' reference ranges. The results showed that data loss due to processing with reference intervals negatively impacted the results. Given the delayed response to COVID-19 in Latin American countries, the datasets for this study were collected from this region. The training dataset for the machine learning model was collected from Ecuador and Peru. External validation was performed using a new, unique dataset. This regional focus highlights the importance of sustainable economic measures to enhance the resilience of health systems in Latin America and facilitate better resource management in future outbreaks.
This paper presents a novel framework for prioritization of green batteries by considering sustainability criteria encompassing socio-economic-environmental-technical aspects. Previous studies on green batteries bring out two questions — what is the importance of sustainable criteria in ranking green batteries and what is the priority of green batteries based on user demand. Since the answers to these questions are subtle, we gain motivation and present an integrated hyperbolic fuzzy based decision framework that methodically calculates experts weights, criteria weights, and customized ranks of batteries. Sustainable criteria considered in this paper are obtained from literature and expert advice as energy density, lifespan/cycle stability, efficiency, chargedischarge rate, availability, recyclability, scalability, temperature sensitivity, cost, and environmental impact of materials. Qualitative rating is interpreted as hyperbolic fuzzy value, followed by variance method is applied for determining experts weights, weighted CRITIC is presented for criteria weights, and query-based rank algorithm is put forward for customized ranking of batteries. Requirement from stakeholders in terms of criteria set is obtained and it is embedded in the CoCoSo formulation for gaining customized ranks. Practicality is demonstrated by considering a case example along with scenario-based query gathering from stakeholders that led following inferences viz., charge/discharge rate and lifespan being top two criteria with Lithium Iron Phosphate and Nickel Metal Hydride batteries being top two batteries. Finally, the pros and cons of the developed framework is presented for aiding policymakers choices.
The digital transformation of cross-border engineering consortia requires robust, scalable, and legally compliant systems for human resource (HR) certification. This study proposes a blockchain-based smart contract framework tailored for cross-border engineering collaborations, addressing critical challenges of multi-jurisdictional compliance, data security, and procedural inefficiency. Leveraging Hyperledger Fabric (a permissioned alliance chain), zk-SNARK (Zero-Knowledge Succinct NonInteractive Argument of Knowledge) protocols, IPFS (InterPlanetary File System) storage, and a dynamic contract engine, the framework integrates sovereign supervision, industry mutual recognition, and enterprise-level validation. Key components include a parameterized qualification matching algorithm, a credit score decay model, and a multilingual smart contract template library. Simulation based on a Southeast Asian high-speed rail project (with prototype validation in two active consortia) demonstrates significant performance improvements against a baseline of traditional paper-based authentication [Formula: see text] diplomatic notarization (widely adopted in cross-border engineering): authentication cycles are shortened by 85.5% (from 47 days to 6.8 days), cost per certification reduced by 93.7% (from $382 to $24), and qualification matching accuracy maintained at 92–94% (defined as consistency with expert manual review). The optimized Raft consensus mechanism (adapted for Hyperledger Fabric) strengthens trust among international nodes. Limitations of zk-SNARK (e.g., trusted setup risks) and scalability tradeoffs are acknowledged, with mitigation strategies proposed. This framework provides a scalable, secure, and legally adaptive model for digitized HR certification, offering substantial operational and regulatory value to cross-border infrastructure initiatives.
Achieving sustainability and resilience in Cold Supply Chain (CDSC) operations has lately attained prominence in healthcare amid climate change, natural catastrophes, and geopolitical crises. This study contributes to the literature by developing a comprehensive framework to establish a sustainable–resilient buyer–supplier relationship to enhance CDSC efficiency. In this context, a Stratified Fuzzy variant of the Best–Worst Method (SFBWM) was proposed to assess the significance of key performance indicators (KPIs) under several future scenarios. Subsequently, the Extended version of the Fuzzy Complex Proportional Assessment (EFCOPRAS) method was presented to assess incumbent suppliers’ performance and classify them into three selection categories: retain, develop, and switch. Later, a Pakistani healthcare case study and extensive sensitivity and comparative analyses determined the practicality and accuracy of the suggested decision support framework. The case study results demonstrate that, regarding suppliers’ capability, the “Timely delivery” KPI gained significant importance, whereas the KPI “Agility” acquired noteworthy importance in terms of suppliers’ desirability.
Smart manufacturing systems integrate technological platforms like cloud computing and computational paradigms to facilitate maintenance, monitoring, and automation. Operations and automation control based on devices or controllers depend on their connectivity and the quantity of machines they manage. This leads to further issues of task failures caused by controller interruptions. This paper introduces a Fault Tolerance Computable (FTC) method utilizing Fractional Learning (FL) to address this issue. The suggested computational approach uses ensemble learning via two fractional ensemble methods: stacking and boosting. During the stacking process, the ratio of the preceding maximum fault tolerance rate is set equal to the current failure rate. This equalization is optimized from median tolerance to a high rate via recurrent training. The training commences in the stacking state utilizing the previously provided controller logs. The boosting state provides an alternative to the existing job allocations and machine schedules derived from the equalization factor obtained in the preceding state. The controller’s machine-control output facilitates swapping at different time intervals. The process is initiated between two consecutive operational states of the industrial controller to optimize task completion rates. Findings: FTC-FL improves task allocation, completion, and tolerance rate by 10.79%, 11.88%, and 9.86%, respectively, with 11.9% fewer failures.
The rapid growth of e-commerce platforms and online user interactions has significantly increased the amount of available user-generated data, which makes recommendation systems essential for delivering personalized content. However, conventional recommendation approaches rely on centralized data, which raises serious concerns about user privacy, data leakage and unauthorized access. To address these challenges, this paper proposes a privacy-preserving personalized recommendation framework that integrates Local Differential Privacy (LDP) with Federated Learning (FL) and transformer-based representation learning. In the proposed model, a Modified Attentional Autoregressive XLNet (MA-XLNet) model is employed to learn contextual representation from user reviews and interaction sequences through dual embedding layers and a soft attention mechanism. To capture both short-term and long-term behavior patterns, a Recurrent Long short Co-ordinate Network (RLCN) is introduced, which integrates LSTM with coordinate attention to generate robust user preference vectors. Furthermore, an LDP-based FL is incorporated to enable secure distributed model training while preventing the leakage of sensitive user information during parameter aggregation. Experimental evaluations are conducted on the Amazon review dataset containing millions of user reviews and product interactions. The proposed model demonstrates superior recommendation performance compared with baseline models, achieving 99.45% accuracy, 99.45% precision, 99.44% recall, and 99.45% [Formula: see text]1-score while also achieving improved ranking performance with low MAE (0.0055) and RMSE (0.074162) values. In addition, the model obtains strong ranking metrics with a Hit Rate (HR) of 99.2% and a Normalized Discounted Cumulative Gain (NDCG) value of 98.56 at higher recommendation thresholds, indicating improved recommendation quality. The results confirm that integrating transformer-based contextual learning with LDP-enabled FL effectively enhances both recommendation accuracy and privacy preservation. The proposed framework provides a scalable and secure recommendation architecture suitable for privacy-sensitive environments such as e-commerce and online platforms.
