Bilgisayar Mühendisliği Bölümü Koleksiyonu

Bu koleksiyon için kalıcı URI

https://hdl.handle.net/20.500.13055/43

Listeleniyor 1 - 20 / 49

Operational outcomes of an AI based streaming speech-to speech translation system in multilingual call centers: A case study
(2026) Bağrıyanık, Selami; Mutlu, Behçet; Kılıç, Gül Damla
Operational outcomes of an AI based streaming speech-to speech translation system in multilingual call centers: A case study
A comparative analysis of divide-and-conquer algorithms in terms of energy consumption and carbon emissions
(Akdeniz Üniversitesi, 2026) Akpınar, Damla; İnce, Ülkü Bensu; İlgın, Onur; Gözüaçık, Necip
This study presents a comparative analysis of four fundamental divide-and-conquer algorithms—MergeSort, QuickSort, conventional matrix multiplication, and Strassen matrix multiplication—in terms of execution time, energy consumption, and carbon emissions. The experiments were conducted in a Python environment using the CodeCarbon library. Random datasets and datasets containing a high proportion of duplicate elements, ranging from 10,000 to 500,000 elements, were used for the sorting algorithms, whereas square matrices ranging in size from 64 × 64 to 256 × 256 were used for the matrix multiplication algorithms. The results demonstrated that QuickSort exhibited lower energy consumption and carbon emissions than MergeSort, particularly for large datasets containing a high proportion of duplicate elements. In the matrix multiplication experiments, despite its theoretical advantage, the Strassen algorithm consumed more energy than the conventional method because of the memory overhead associated with its recursive structure. These findings indicate that algorithm selection should consider not only theoretical time complexity but also energy efficiency and environmental impact, emphasizing the importance of sustainable software development.
Multidimensional diversity in video recommender systems: A holistic framework of literature gaps and future directions
(Elsevier, 2026) Dönmez, İlknur; Bulut, Faruk
Video recommender systems shape content discovery, user engagement, and information access across digital platforms such as YouTube, Netflix, and TikTok. Although recent advances in deep learning, graph-based models, and multimodal recommendation have improved personalization, key challenges remain, including limited di versity metrics, weak geometric analysis of embedding spaces, inadequate multimodal datasets, and insufficient integration of fairness and transparency. This review synthesizes recent research on diversity-aware video recommender systems through a structured literature review. The screening process identified 572 records from six scholarly sources; after title–abstract screening and full-text eligibility assessment, 101 studies were included in the qualitative synthesis. The findings show that diversity is often treated as a narrow, one-dimensional objective, while broader system-level, user-centered, and societal perspectives remain underdeveloped. The re view therefore argues for multidimensional and human-centered recommendation frameworks that incorporate geometric diversity modeling, multi-objective optimization, long-term user satisfaction, and user-controllable design. The study provides a conceptual basis for developing more robust, fair, transparent, and sustainable video recommender systems.
Problem formulation outweighs loss-function choice in deep-learning cervical vertebral maturation staging: A multi-seed study on an imbalanced public benchmark
(MDPI Publishing, 2026) Tokatlı, Nazlı
(1) Background: Cervical vertebral maturation (CVM) staging on lateral cephalograms informs the timing of orthodontic treatment, and automated deep-learning approaches are widely studied. A recurring but under-examined obstacle is that public CVM datasets are severely imbalanced across the six stages. Using a public dataset and a rigorous eight-seed protocol, we quantify how strongly imbalance constrains performance, whether common imbalance-handling and ordinal-aware loss functions overcome it, and whether a clin ically motivated three-group reformulation is more tractable than the native six-stage task. (2) Methods: Using the public Aariz dataset (1000 lateral cephalograms with expert CVM-stage labels; 700/150/150 train/validation/test, one radiograph per patient), we trained an ImageNet-pretrained ResNet-18 under four strategies: unweighted cross-entropy (baseline), class-weighted cross-entropy with balanced sampling, focal loss, and a rank consistent ordinal (CORAL) head. Each configuration was trained with eight random seeds. Both tasks were evaluated over this multi-seed protocol, and sensitivity analyses examined the ordinal model’s sampling scheme and the focal-loss focusing parameter. Performance was assessed by accuracy, macro-F1, mean absolute error in stages (MAE), and quadratic weighted kappa (QWK), and compared across strategies with the Kruskal–Wallis test; sec ondary pairwise contrasts were Holm-corrected, and Wilson binomial confidence intervals (accuracy), seed-level confidence intervals (macro-F1), and analytically derived major ity and frequency-weighted random classifiers were added as trivial reference baselines. Both the native six-stage task and a three-group scheme (pre-peak = CS1-3, peak = CS4, post-peak = CS5-6) were evaluated. (3) Results: On the six-stage task the extreme imbal ance (CS1 n = 18 vs. CS5 n = 311 in training) produced poor minority-stage recognition (macro-F1 ≈ 0.22–0.25), with the rarest stages essentially unclassified even after imbalance handling. Reframing into three clinically meaningful groups roughly doubled macro-F1 (to ≈0.47–0.51) across all strategies. Critically, in the three-group setting no loss function significantly outperformed the others: across eight seeds the Kruskal–Wallis test was non significant for every metric (macro-F1 p = 0.09, QWK p = 0.05, MAE p = 0.07, accuracy p = 0.41), and the unweighted baseline matched the best specialized loss on macro-F1 (0.507 ± 0.045 vs. ordinal 0.505 ± 0.044). In secondary pairwise comparisons the ordinal loss nominally exceeded the weighted and focal losses (uncorrected p = 0.04 and p = 0.01), but after Holm correction only the comparison with focal loss remained significant(adjusted p =0.03), and no pairwise difference survived a conservative six-comparison Bonferroni bound; focal loss was the least stable. All strategies clearly exceeded trivial baselines on macro-F1 (majority classifier 0.259; frequency-weighted random ≈0.333), although the majority classifier’s accuracy (0.633) exceeded that of every trained model. (4) Conclusions: On this imbalanced public dataset, the dominant factor associated with CVM classification performance was the problem formulation rather than the loss function: a clinically grounded three-group scheme was far more tractable than six-stage classification, whereas specialized imbalance and ordinal losses did not reliably outperform a simple baseline. These findings, established under a fixed backbone, training budget, and a single public dataset, caution against assuming that loss-level fixes resolve CVM imbalance, and high light data scale and problem formulation as the primary levers. Multi-dataset external validation is a necessary next step.
Hepdecide: A standards-oriented, explainable clinical decision support prototype for the chronic liver-disease spectrum—design, guideline-grounded reporting, and a leakage-controlled retrospective evaluation
(MDPI Publishing, 2026) Tokatlı, Nazlı
Background/Objectives: Specialist hepatology capacity is small relative to the population needing assessment, so most decisions about whom to investigate for chronic liver disease fall to non-specialists guided by static laboratory indices reported in isolation. This study presents HepDecide, a standards-oriented decision support prototype consolidating stage specific risk estimation, explainability and guideline-traceable reporting into one physician facing service, and reports a leakage-controlled retrospective evaluation of its predictive layer. Methods: HepDecide combines stage-specific tabular classifiers, an explainability layer based on SHapley Additive exPlanations (SHAP), and retrieval-augmented genera tion (RAG) anchored to EASL and AASLD guidelines; outputs are expressed as HL7 FHIR R4 resources, with inputs mapped to LOINC and diseases to ICD-10. Three endpoints were modelled on two publicly available, de-identified cohorts: steatohepatitis versus simple steatosis and histological cirrhosis (F4) in a biopsy-characterised MASLD cohort (n = 605), and one-year survival among patients with hepatocellular carcinoma (HCC; n = 165). Because patient-level predictions from the original development runs were not retained, the primary evaluation is a separate re-analysis with a locked patient-wise test partition, imputation, scaling and resampling fitted within training folds only, 25 classifier–imbalance configurations per endpoint, training-only calibration, and calibration, precision–recall, decision-curve and SHAP analyses. Results: On the locked test partitions, ROC-AUC was 0.67 (95% CI 0.50–0.84) for steatohepatitis (n = 121), 0.81 (0.52–1.00) for cirrhosis (n = 121; five events) and 0.76 (0.56–0.92) for one-year HCC survival (n = 33), with Brier scores of 0.096, 0.028 and 0.199. On the same cirrhosis test partition FIB-4 reached an ROC-AUC of 0.92 (0.82–0.99). The five highest-ranked SHAP features accounted for 58–73% of total attribution. Development-phase estimates of 0.90, 0.98 and 0.92 are reported separately as archived values; the cirrhosis estimate is documented in a retained training output, although its original input tables and individual test probabilities are unavailable. No quantitative evaluation of the language-generation layer is claimed, because the corre sponding records were not retained. Conclusions: Under leakage-controlled retrospective conditions the predictive layer showed moderate discrimination with wide confidence intervals and did not outperform FIB-4 for cirrhosis. The value of the prototype lies in its standards-oriented, explainable and guideline-grounded design and in a fully docu mented, reproducible evaluation pipeline; transportability, the text-generation layer and clinical usefulness remain to be established prospectively. HepDecide is intended as a clinician-supervised triage and reporting prototype, not a clinically ready system.
Air ticket sales forecasting for B2B travel agencies
(Emerald Publishing, 2026) Uzun-Per, Meryem; Akkaya, Ümit Murat; Kalkan, Habil
Purpose – This study addresses the air ticket sales forecasting problem for business-to-business (B2B) travel agencies, a largely unexplored area compared to airlines and business-to-consumer agencies. To meet guarantee deposit requirements and manage financial risk exposure, B2B agencies require accurate demand estimation for budget planning and revenue management. The study proposes a hybrid deep learning forecasting model integrating recurrent neural network (RNN), long short-time memory (LSTM) and bidirectional long short-term memory (BiLSTM) architectures to capture both short-term fluctuations and long-term temporal dependencies in ticket sales data. Design/methodology/approach – Approximately eight years (from 3 August 2014 to 2 June 2022) of daily search and ticket sales data from a B2B travel consolidator agency were combined with external variables, including holidays, weather conditions, exchange rates, gross domestic product per capita and COVID-19 cases. Two datasets (with and without pandemic data) were constructed. After preprocessing and Random Forest feature selection, seasonal autoregressive integrated moving average (SARIMA), Prophet, RNN, LSTM and BiLSTM models were implemented and compared with the proposed hybrid architecture using a sliding window forecasting framework. Performance was evaluated using mean absolute percentage error (MAPE) and root mean squared error (RMSE). Findings – The proposed hybrid model achieved the highest forecasting accuracy across both datasets, outperforming statistical and individual deep learning models. It obtained MAPE values of 5.83% and 6.16% and RMSE values of 0.0379 and 0.0411, respectively. BiLSTM ranked as the second-best method, while SARIMA and Prophet showed substantially lower accuracy. Results also indicated that excluding pandemic variables did not significantly reduce predictive performance, but still, pandemic-related volatility increased the error rate of most of the methods and increased forecasting difficulty. Originality/value – This research is the first to examine ticket sales forecasting specifically for B2B travel agencies. It introduces a novel hybrid sequential deep learning architecture and evaluates the role of pandemic and macroeconomic variables in forecasting performance. The findings demonstrate the effectiveness of hybrid recurrent models for complex real-world demand forecasting problems.
Dynamic virtual credit cards in OTA B2B payments: An industrial case study
(Euroasia Journal of Social Sciences & Humanities, 2026) Bağrıyanık, Selami; Mutlu, Behçet
Introduction and Purpose: Online travel agencies (OTAs) rely on secure B2B payment integrations. Conventional models often use a "stored-card exposure" approach, storing customer cards reversibly without usage limits, creating severe security vulnerabilities. This study evaluates an architectural transformation to a centralized Virtual Credit Card (VCC) pipeline, WingieVCC, designed for transaction isolation and automated balance management within a high-volume OTA ecosystem, specifically targeting the hospitality sector. Materials and Methods: An industrial case study methodology evaluated this transformation over a one-year deployment period (Q2 2024 to Q1 2025). The system dynamically generates transaction isolated, disposable VCCs with strict, currency-based limits. It integrates multi-currency balance tracking, automated threshold routing, and an idempotent API framework with retry mechanisms to handle network timeouts. During deployment, the system utilized shadow traffic and staged rollout strategies to ensure seamless integration. The evaluation focused on architectural resilience, security posture, and operational improvements. Results: The architectural shift yielded profound qualitative improvements. By migrating from a limitless, stored-card model to a transaction-isolated, disposable VCC architecture, card data exposure was structurally eliminated. Defining strict currency-based limits for each transaction guaranteed that even in breach scenarios, unauthorized usage is technically impossible; unused or failed cards are automatically invalidated. Operationally, centralized balance management significantly reduced failed transactions caused by insufficient personal card limits, minimizing the need for manual interventions. Performance-wise, the system successfully met a Service Level Agreement (SLA) target of < 600 ms for transaction response times. Replacing traditional, stored-card infrastructures with a dynamic, transaction-isolated VCC pipeline fundamentally mitigates critical security risks in OTA platforms. This architectural transformation demonstrates that integrating strict transaction limits, shadow traffic deployment, and idempotent APIs maximizes transaction reliability and operational efficiency. Ultimately, this isolated architecture offers a highly scalable and secure blueprint for advancing fintech integrations within the global tourism sector.
Machine learning-guided green ultrasound processing enhances polyphenols, antioxidant capacity, and in vitro bioaccessibility of black fig vinegar
(Frontiers Media S. A., 2026) Çakmak Sancar, Burcu; Aktaran Bala, Deniz; Demirel, Selinay; Gürfidan, Remzi; Yıkmış, Seydi; Kilim, Oğuzhan; Açıkgözoğlu, Enes; Şimşek, Mehmet Ali; Tokatlı Demirok, Nazan; Ibrahim, Marwa Ezz El-Din; Tokatlı, Nazlı
Black fig vinegar is a functional fermented product rich in phenolic compounds and antioxidants; however, conventional thermal pasteurization may reduce its bioactive composition and functional quality. Although ultrasound processing has emerged as a promising non-thermal alternative, its optimization, predictive modeling, and influence on the in vitro bioaccessibility of black fig vinegar have not yet been comprehensively investigated. Therefore, this study aimed to optimize ultrasound processing conditions and evaluate their effects on phenolic composition, antioxidant capacity, and in vitro bioaccessibility of traditionally produced black fig vinegar using response surface methodology (RSM) integrated with ensemble machine learning algorithms. Untreated, pasteurized, and ultrasound-treated vinegars were compared, while ultrasound processing variables (processing time and amplitude) were optimized through RSM. The experimental dataset was subsequently expanded using data augmentation, and Extra Trees, CatBoost, Gradient Boosting, and AdaBoost models were developed to predict total phenolic content (TPC) and ferric reducing antioxidant power (FRAP). Among the tested algorithms, the Extra Trees model achieved the best predictive performance, with coefficients of determination (R2) exceeding 0.998 and prediction accuracies above 99%. Under the optimized ultrasound conditions (8 min and 58% amplitude), TPC (79.77 ± 3.69 mg GAE/100 mL), FRAP (8.87 ± 0.24 mmol TE/L), and total flavonoid content were significantly higher than those of untreated and pasteurized samples (p < 0.05). Ultrasound processing also significantly increased caffeic acid, catechin hydrate, chrysin, vanillin, p-coumaric acid, o-coumaric acid, and trans-ferulic acid concentrations. During simulated gastrointestinal digestion, all samples exhibited reductions in bioactive compounds; however, ultrasound-treated vinegar consistently retained higher TPC, total flavonoid content, and FRAP values, with bioaccessibility remaining approximately 30%–32%. Pearson correlation and principal component analyses further identified caffeic acid, trans-ferulic acid, rutin, and o-coumaric acid as the principal contributors to antioxidant capacity. This study integrates ultrasound optimization, RSM, data enhancement, community machine learning, phenolic profiling, and in vitro bioavailability assessment for black fig vinegar, presenting a robust and sustainable strategy for process optimization and the development of high-quality functional vinegar products.
Multi-class network intrusion detection: Machine and deep learning benchmark with live stream deployment
(IEEE, 2026) Danışmaz, Betül; Bıyık, Mustafa Emre; Gözüaçık, Necip
The rapid rise of advanced cyber threats is exposing the limitations of traditional, signature-based network intrusion detection systems. While machine learning and deep learning models offer proactive defense mechanisms, current research is often compromised due to temporal data leakage, impractical threat classifications, and significant discrepancies between offline evaluation and online deployment. This study presents a comparative evaluation of traditional machine learning and deep learning architectures integrated into a fully functional real-time inference pipeline. Additionally, we propose an operationally motivated three-class classification encompassing Benign, Volumetric, and Semantic traffic to contextualize incident response. When the trained XGBoost, Random Forest, Decision Tree, LSTM, and BiLSTM models were evaluated and compared under the same constraints, the BiLSTM model achieved the highest Macro F1 score at 97.72%. However, XGBoost closely followed with a Macro F1 of 95.87%, while achieving 15.5 times faster inference. To translate these findings into practice, a Kafka-based live system architecture was designed that integrates real-time stream inference, zero-downtime model switching, and automated threat response capabilities.
Driver fatigue estimation using multimodal deep learning with eeg and eog signals
(IEEE, 2026) Demir, Șevval; Yapar, Rümeysa; Aldaş, Ahsen Nur; Gözüaçık, Necip
Driver fatigue is a critical condition that increases accident risk by causing loss of concentration and perceptual limitations. Therefore, researchers have focused on developing systems aimed at the early detection of driver fatigue. In this study, driver fatigue detection was targeted using Differential Entropy–Linear Dynamic System (DE-LDS) based Electroencephalography (EEG) features and Electrooculography (EOG) features provided by the SEED-VIG dataset. The feature pool employed in the study was enriched through feature engineering methods such as band ratios, temporal delta components, hemispheric asymmetry, and moving standard deviation. Moreover, by eliminating the need for raw signal processing, the focus was directed toward the temporal variation patterns of meaningful features. Experimental results demonstrated that predictions based solely on EEG data remained limited, whereas the inclusion of EOG data significantly improved performance.These findings reveal that EEG and EOG signals provide complementary information in the prediction of driver fatigue. The proposed multimodal approach exhibited greater robustness against inter-subject variability compared to unimodal systems and offered a more generalizable monitoring framework.
Thermosonication improves the bioactive properties and in vitro bioaccessibility of broccoli juice through integrated RSM and metaheuristic optimization
(Frontiers Media S. A., 2026) Arpa Zemzemoğlu, Tuba Eda; Tokatlı Demirok, Nazan; Türkol, Melikenur; Gürfidan, Remzi; Kilim, Oğuzhan; Tokatlı, Nazlı; Yıkmış, Seydi; Mohamed Ahmed, Isam A.; Aljobair, Moneera; Yinanç, Abdullah
Broccoli juice is highly nutritious, containing phenolic compounds, chlorophyll derivatives, antioxidants, and other biologically active phytochemicals; however, conventional thermal processing may reduce their stability and bioaccessibility. The aim of this study was to analyse the impact of thermosonication on the following properties of broccoli juice: bioactivity, antioxidant capacity, phenolic profile and in vitro bioaccessibility. Thermosonication conditions were optimized using Response Surface Methodology (RSM) integrated with the Grey Wolf Optimizer (GWO) and Particle Swarm Optimization (PSO) algorithms. Broccoli juice samples were subjected to conventional pasteurization and thermosonication under different processing conditions, and total chlorophyll, ascorbic acid (AA), ferric reducing antioxidant power (FRAP), total phenolic content (TPC), and individual phenolic compounds were evaluated. Digestive stability and post-digestion bioaccessibility were also assessed using a standardized in vitro gastrointestinal digestion model. The developed quadratic models exhibited high statistical significance for TPC and FRAP responses (p < 0.001), with coefficients of determination (R²) of 0.9949 and 0.9853, respectively. The optimum thermosonication conditions were identified as 40 °C, 6.75 min, and 64.04% amplitude, under which experimentally validated FRAP and TPC values reached 12.86 mmol TE/L and 149.77 mg GAE/100 mL, respectively. HPLC-DAD analysis showed that gallic acid and naringin were the predominant phenolic compounds, and that thermosonication promoted greater phenolic retention than conventional pasteurization. In vitro digestion results demonstrated that thermosonicated broccoli juice maintained significantly higher levels of TPC, FRAP, chlorophyll, and AA throughout gastrointestinal digestion, with higher TPC recovery (30.45%) than pasteurized samples (26.22%), indicating improved post-digestion bioaccessibility of phenolic compounds. Similar improvements were observed for antioxidant capacity and chlorophyll retention. Furthermore, PSO and GWO optimization results showed strong agreement with RSM predictions, confirming the robustness of the proposed optimization strategy. Overall, thermosonication was shown to effectively and sustainably enhance the functional quality, antioxidant properties, phenolic stability, and digestive bioaccessibility of broccoli juice.
Mathematical and numerical modeling of drilling system dynamics using CF fractional differentiation
(AIMS Press, 2026) Erdoğan, Mahir Ceylan
In this work, I propose a Caputo–Fabrizio fractional mathematical model of a drilling system (CFFMMDS) to investigate the coupled dynamics of an induction motor-driven drilling assembly within a unified fractional-order framework. The governing equations were formulated using a nonsingular Caputo–Fabrizio fractional derivative with an exponential memory kernel to characterize the temporal dependencies among the system state variables. The resulting fractional-order system was solved using the Caputo–Fabrizio q-Elzaki homotopy analysis transform method (CFq-EHATM) to derive semi-analytical solution series and to conduct a numerical investigation of the system response under varying fractional orders. The results indicated that a decrease in the fractional order 𝜇 results in attenuated growth and decay rates. This trend is consistent with the intrinsic memory structure of the Caputo–Fabrizio operator, which distributes the influence of past states over time. The proposed framework enables the analysis of the temporal evolution of coupled state variables in dynamic systems exhibiting history-dependent behavior.
An NLP-driven framework for automated radiology–pathology concordance assessment in breast biopsy
(MDPI Publishing, 2026) Esmerer, Emel; Nazlı, Mehmet Ali; Uzun-Per, Meryem; Gümüş Değidiben, Melike; Söyleyici, Merve; Tahir, Eren; Bal, Mert
Background/Objectives: To develop and assess the feasibility of a natural language processing (NLP) framework for automated assessment of radiology-pathology concordance in breast biopsy using machine learning-based analysis of unstructured reports. Methods: This retrospective study included 766 paired radiology and pathology reports from ultrasound- or mammography-guided breast biopsies (August 2020-May 2024). Reports underwent translation, normalization, tokenization, lemmatization, and synonym expansion, followed by structured encoding of BI-RADS and pathology categories. Three models were trained: a Decision Tree, a LightGBM classifier, and a fine-tuned BioBERT model. Concordance labels were defined by multidisciplinary consensus. Performance metrics included accuracy, sensitivity, specificity, F1-score, area under the curve (AUC), and Cohen's kappa. SHapley Additive exPlanations (SHAP) analysis was used to identify influential features. Results: Among 766 cases, 707 (92.3%) were concordant and 59 (7.7%) were initially discordant. After excluding B3 lesions (n = 46), 13 true discordant cases remained (1.7%). Including B3 lesions increased clinically non-concordant or indeterminate cases from 1.7% to 7.7%, indicating that the apparent performance of the models is likely sensitive to case definition and dataset composition. BI-RADS 4a was the most common category (31.3%), and benign pathology (B2) accounted for 64.4% of biopsies. Within this dataset, LightGBM yielded the highest apparent AUC (0.999) (however, given the extremely small number of true discordant cases, this estimate is likely unstable and should be interpreted with caution), while BioBERT showed the strongest agreement with expert consensus (κ = 0.89). SHAP analysis identified clinically meaningful terms such as calcification, hypoechoic, ductal, and carcinoma as key contributors to model predictions. Given the very limited number of true discordant cases, these performance estimates are likely unstable and should be regarded as preliminary, requiring validation in larger, multi-center cohorts. Conclusions: This study presents a proof-of-concept NLP-based framework for radiology-pathology concordance assessment. The models showed promising performance in identifying potentially discordant cases; however, given the limited number of true discordant samples, these findings should be considered preliminary and require further validation in larger, multi-center datasets before clinical implementation.
Comparative benchmarking of 2d lidar slam algorithms with ros 2 on raspberry pi 5
(IEEE, 2026) Şahin, Ulaş; Can, Göktürk; Altıok, Ezgi; Çavdar, İbrahim; Gözüaçık, Necip
Mobile robotics increasingly relies on SLAM for robust autonomous navigation. While many algorithms exist, systematic comparisons within the ROS 2 framework under real-world conditions remain limited. This study addresses this gap by benchmarking three widely used 2D LiDAR-based methods—GMapping, Hector SLAM, and Cartographer—on a wheeled mobile robot. Using both simulation and on-device experiments, we evaluate mapping accuracy, localization quality, and computational efficiency. Results show that Cartographer achieves the highest accuracy in structured environments, Hector SLAM demonstrates robustness without odometry, and GMapping performs reliably only in small-scale settings. These findings highlight trade-offs relevant to embedded deployment. The main contributions are: (i) a reproducible evaluation pipeline on ROS 2, (ii) quantitative analysis of accuracy versus resource usage on Raspberry Pi 5, and (iii) practical guidelines for algorithm selection in autonomous systems. This work advances the understanding of ROS 2-based SLAM and supports informed deployment in robotics applications.
Meta-learning analysis of deep neural network architectures on diverse numeric datasets via geometric complexity descriptors
(Wiley, 2026) Bulut, Faruk; Dönmez, İknur
Meta-learning techniques aim to predict the most suitable learning algorithm for a given dataset based on its intrinsic structural characteristics. These techniques provide a robust framework for understanding algorithmic behavior across diverse data dis tributions and attributes. Although these state-of-the-art models (CNNs and transformers) are widely applied in various machine learning tasks, their use on numerical datasets remains underexplored due to the complexity of their internal structures. This study aims not only to predict the performance of two black-box deep learning models on static datasets but also to conduct a behavioral analysis in order to identify which meta-features most strongly infuence their outcomes. It seems unclear which specifc attributes of a dataset positively or negatively afect the performance of these deep learning models. To bridge this gap, we constructed a meta dataset consisting of 296 datasets, each characterized by 20 meta-features describing the dataset’s statistical, geometric, and structural properties. The analysis identifes which intrinsic dataset properties infuence model accuracy, without relying on raw data or hyperparameter tuning. Results show that both models perform best on datasets with high feature discriminability, as captured by meta-features such as maximum feature efciency, collective feature efciency, and directional separability. In contrast, performance declines with increasing class boundary complexity and nonlinearity, refected in features like class separability measures and the linear classifer nonlinearity metric. While CNNs are more sensitive to local geometric complexity, transformers respond more strongly to global statistical measures such as mutual information and entropy, highlighting their distinct inductive biases. The proposed meta-model accurately predicts the performance of both architectures on unseen datasets (0.96 correlation coefcient, 0.019 MAE, and 0.025 RMSE for CNNs; 0.92 correlation coefcient, 0.027 MAE, and 0.036 RMSE for transformers), enabling performance estimation without costly training. These fndings emphasize the importance of aligning model architecture with dataset geometry and structure. Additionally, the framework supports more interpretable, efcient, and sustainable deep learning model selection in structured data settings.
A comprehensive review on the use of artificial intelligence, internet of things, sensors, and green energy in non-invasive agricultural techniques
(FRUCT, 2025) Serdaroğlu, Kemal Çağrı; Tokatlı, Nazlı
Feeding a burgeoning global population amid cli mate change and dwindling resources presents a profound chal lenge for agriculture. This paper examines ”smart agriculture” (Agriculture 4.0) as a pivotal solution, integrating technologies like IoT, AI, and robotics to cultivate data-driven, efficient, and sustainable farming. We emphasize the growing effectiveness of multi-modal data fusion—combining diverse sensor inputs—for improved pest detection, water management, and yield predic tion. A critical shift towards decentralized edge intelligence is also explored, facilitating real-time, on-farm decisions and overcoming connectivity hurdles. While acknowledging that successful implementations are highly context-specific and that synthetic data can address scarcity, we also confront persistent obstacles: high adoption costs, the digital divide, unreliable rural connectivity, and cybersecurity risks. Ultimately, realizing smart agriculture’s full potential—a more resilient and productive global food system—requires sustained investment in affordable sensors, robust and explainable AI, and autonomous robotics to translate data insights into actionable field-level strategies.
ThermoMicrowave-sonication improves the stability and digestive bioaccessibility of phenolic compounds in parsley juice
(Elsevier, 2026) Yıkmış, Seydi; Tokatlı Demirok, Nazan; Duman Altan, Aylin; Paçal, İshak; Türkol, Melikenur; Tokatlı, Nazlı; Paçal, Nurettin; Abdi, Gholamreza; Aadil, Rana Muhammad
These are indications of the effects of ThermoMicrowave Sonication (TMS) on the bioactive compounds of parsley (Petroselinum crispum) juice and their bioaccessibility during in vitro digestion. Total phenolic content (TPC), iron-reducing antioxidant power (FRAP), chlorophyll, and ascorbic acid levels were measured in TMS treated and pasteurized samples. TMS minimized the loss of heat-sensitive proteins and significantly increased the phenolic content and antioxidant structure (p < 0.05). By following simulated oral, gastric, and intestinal digestion, TPC, chlorophyll, and FRAP levels were better in TMS samples than in controls or pasteurized samples. The highest recoverable levels were observed in the intestinal phase, highlighting the role of TMS in supporting functional quality after digestion. Prediction models using linear regression and LASSO showed strong accuracy (R2 > 0.99) for antioxidant capacity. Overall, TMS offers a promising, environmentally friendly, and industrially applicable tool for preserving and ensuring bioaccessibility of bioactive images in parsley juice and valuable information for functional electrical development. Chemical compounds: Gallic acid (PubChem CID:370); flovone (PubChem CID: 10680); vanillic acid (PubChem CID: 8468); rutin (PubChem CID: 5280805); naringin (PubChem CID: 442428); p- coumaric acid (PubChem CID: 637542); o- coumaric acid (PubChem CID: 637540); quercetin (PubChem CD: 5280459); alizarin (PubChem CD: 6293).
SFNN: A secure and diverse recommender system through graph neural network and regularized variational autoencoder
(Elsevier, 2025) Bahi, Abderaouf; Gasmi, Ibtissem; Bentrad, Sassi; Azizi, Mohamed Walid; Khantouchi, Ramzi; Uzun-Per, Meryem
Recommender systems are frequently improved to filter information and provide users with the most relevant items. However, they face limitations in balancing appropriate and diverse recommendations while ensuring the security and integrity of user data. A new recommender system based on secure fusion neural network is pre sented in this paper. It guarantees data integrity and confidentiality while balancing accuracy and diversity. It integrates a graph neural network that models user-item interactions to improve accuracy, with a regularized variational autoencoder whose evidence lower bound loss function is enhanced by a diversity-promoting regu larization term that favors latent-space dispersion, thereby improving recommendation diversity. To optimize the combination of the two neural networks scores, an adaptive fusion mechanism is introduced to generate final predictions that consider diverse user preferences while maintaining relevance. Furthermore, our approach uses blockchain technology to encrypt and secure data storage, ensuring the integrity and confidentiality of users’ data. The experiments conducted on three datasets show that the proposed model can achieve an accuracy of 78.13 % with an intra-list diversity of 46.82 % for Retail Rocket dataset, an accuracy of 82.44 % with an intra-list diversity of 37.78 % for clothing dataset, and an accuracy of 86.16 % with an intra-list diversity of 47.65 % for MovieLens-1 M dataset.
A comparative study of deep learning models for automated liver and tumor segmentation in 2d contrast-enhanced MRI images
(IEEE, 2025) Tokatlı, Nazlı; Bilmez, Yakuphan; Bayram, Mücahit; Bayır, Beyzanur; Özalkan, Helin; Tekin, Zeynep; Örmeci, Necati; Altun, Halis
This paper presents a comprehensive investigation into deep learning techniques for the automated segmentation of the liver and tumors from 2D abdominal contrast-enhanced Magnetic Resonance Imaging (MRI) slices. Addressing a significant challenge in medical image analysis, our study leverages the public ATLAS dataset [1], using a selection of 60 3D abdominal MRI scans, from which we extracted approximately 3,750 2D slices for model training and evaluation. The core objective was the precise identification and delineation of both the liver organ and any intrahepatic lesions. A comparative analysis was conducted on three U-Net-based architectures: the standard Attention U-Net model incorporating EfficientNet-b3 and CBAM but without Focal Loss, the Attention U-Net model with integrated Focal Loss, and the ResNet34-Based U-Net model. To optimize performance, we explored the efficacy of different loss functions, namely DiceLoss and a hybrid DiceLoss with Focalcoss. Our findings are promising: Among the evaluated models, the ResNet34-Based U-Net demonstrated the highest performance with a Dice score of 91.36% and an IoU score of 89.52%. It was followed by the Attention U-Net with Focal Loss, which achieved 86.41% Dice and 81.61% IoU scores, and the standard Attention U-Net, which obtained 85.93% Dice and 81.19% IoU scores. These results underscore the significant potential of our 2D-based methodology to enhance the precision and efficiency of liver and tumor detection from abdominal scans, offering a valuable tool to support clinicians in early diagnosis and to alleviate their workload.
An AI-powered mobile application for aroid identification and interactive learning: Enhancing pharmacognosy education through deep learning and NLP
(IEEE, 2025) Tokatlı, Nazlı; Bilmez, Yakuphan; Kılıç, Yusuf; Alpınar, Abdülkerim
Aroid plants (Araceae family), recognized for their distinct inflorescence, possess significant botanical, pharmaceutical, and practical importance due to their content of both beneficial compounds and toxins such as calcium oxalate crystals. Accurate identification of these species is particularly crucial in pharmacy education; however, morphological similarities among Aroid species often lead to confusion among students. This paper presents a deep learning-based mobile application designed to support both plant identification and interactive learning. The solution leverages EfficientNet and Convolutional Neural Network (CNN) architectures, achieving up to 96 % accuracy in classifying Aroid species. The visual classification model, trained on a comprehensive dataset, is deployed via a RESTful API and integrated within a Flutter-based mobile application. In addition, the app incorporates a Natural Language Processing (NLP)-powered chatbot to address user inquiries regarding plant characteristics and care. While technical evaluations demonstrate robust model performance, a comprehensive user evaluation aimed at assessing the system's educational value, usability, and chatbot interaction is planned as future work. This study underscores the potential of AI-driven mobile solutions in advancing pharmacognosy education, with future developments aimed at expanding the app's botanical scope and enhancing user engagement based on forthcoming survey results.

Güncel Gönderiler