Industrial spray drying is a complex nonlinear process characterized by strong interactions among thermal, mechanical, environmental, and temporal variables, making accurate prediction and operational optimization challenging. This study proposes a hybrid machine learning–evolutionary optimization framework for predictive modeling and intelligent operational optimization in industrial spray drying systems. The framework integrates Random Forest Regression (RFR) for nonlinear process prediction with a Genetic Algorithm (GA) for adaptive exploration of optimal operating conditions under industrial constraints. The RFR model achieved strong predictive performance, with a coefficient of determination of R² = 0.9144 and stable generalization performance (cross-validation R² = 0.8427 ± 0.2027), demonstrating robustness under dynamic industrial variability. Feature importance analysis revealed gas flow and temporal dynamics as the dominant process drivers, enhancing interpretability and supporting operational decision-making. The GA-based optimization identified operating configurations that improved process efficiency by approximately 5.91% compared to baseline conditions while maintaining operational feasibility. The results demonstrate that integrating ensemble learning with evolutionary optimization provides both high predictive accuracy and actionable optimization capability for complex industrial systems. Furthermore, the proposed surrogate-assisted framework reduces computational dependency on physical experimentation by enabling rapid evaluation of candidate operating conditions through data-driven intelligence. The study contributes a deployable data-driven optimization architecture for intelligent and energy-efficient manufacturing operations within Industry 4.0 environments.
Digital grocery shopping has shifted consumer patterns toward multi- item purchasing. While bundle recommendation systems address this, existing models relying on product ID co-occurrence fail to capture the mixed shopping intentions inherent in consumer baskets. To address these limitations, we propose Intent-aware Bundle Recommendation (IABR), a framework shifting from structural matching to semantic intent reasoning. IABR utilizes Large Language Models to decompose baskets into coherent sub-packages. Subsequently, we fine- tuned Gemma-3-4B using Parameter-Efficient Fine-Tuning to generate narrative intent descriptions regarding short-term shopping missions and long-term sustainable user lifestyles. These intents are encoded via Sentence Transformers for semantic retrieval. Extensive testing on the Instacart dataset demonstrates IABR’s significance against baselines like BGCN. Our IABR method achieved a Recall@20 of 28.15% while improving diversity scores by 12% (p < 0.05). This validates that generative semantic modeling enables accurate next- bundle predictions, effectively balancing precision with thematic variation and personalization.
Skewed class distribution continues to be one of the central obstacles in binary classification, since a learning model tends to lean toward the dominant class and consequently overlooks observations belonging to the under-represented class. The purpose of this research is to examine how the choice of distance measure inside SMOTE, specifically Euclidean, Manhattan, Chebyshev, and Hamming, affects predictive quality on imbalanced binary data. Ten publicly available binary datasets drawn from the KEEL repository, whose imbalance ratios span from 1.86 up to 15.80, were used in the experiment. Every dataset was preprocessed and partitioned into 80% for training and 20% for testing; oversampling with SMOTE was carried out on the training portion only, after which four learners, namely Naive Bayes, Decision Tree, Logistic Regression, and k-Nearest Neighbor, were assessed. Model quality was judged through the Matthews Correlation Coefficient (MCC) together with the G-Mean, as these two indicators describe imbalanced performance more faithfully than plain accuracy. The comparison revealed that pairing Euclidean-based SMOTE with Logistic Regression yielded the strongest average scores (MCC = 0.72; G-Mean = 0.79); Manhattan-based SMOTE reached its top MCC again with Logistic Regression (MCC = 0.68) and its top G-Mean with the Decision Tree (G-Mean = 0.79); Chebyshev-based SMOTE delivered the best overall combination together with the Decision Tree (MCC = 0.74; G-Mean = 0.84); and Hamming-based SMOTE performed best alongside Logistic Regression (MCC = 0.73; G-Mean = 0.81). Taken together, these outcomes suggest that the distance function chosen within SMOTE shapes the quality of the generated synthetic points and, in turn, the behavior of the trained classifier.
Machine learning models deployed in sensor-based production environments are prone to performance degradation due to evolving data distributions, commonly known as data drift. In gas sensor systems, such drift often arises from environmental variability and sensor aging, which can significantly reduce predictive reliability if left unaddressed. This study presents an integrated Machine Learning Operations (MLOps) framework that combines performance monitoring, distribution-based drift detection, and adaptive retraining within a unified production pipeline. Experiments are conducted using the Gas Sensor Array Drift Dataset, organized into sequential batches to emulate real-world deployment conditions. Data drift is quantified using Population Stability Index (PSI) and Kullback–Leibler Divergence (KL), which serve as decision thresholds for triggering retraining. The proposed adaptive retraining strategy is systematically compared with baseline (no retraining) and periodic retraining approaches. The results indicate that the adaptive strategy maintains more stable performance across data batches while minimizing unnecessary retraining processes. Additionally, the use of containerization and experiment tracking ensures reproducibility and supports full lifecycle traceability. Overall, this study demonstrates that integrating MLOps practices into sensor-based machine learning systems is essential for improving robustness and ensuring long-term operational sustainability in dynamic environments.
Meta-learning is an approach designed to address data limitations in few-shot learning scenarios. The performance of meta-learning is influenced by the quality of the initial weights used during the meta-training process. Initial weights derived from a relevant domain have the potential to produce more informative feature representations, thereby enabling the adaptation process to new tasks to proceed more effectively. This study analyzes the impact of domain-specific pretrained initialization on classification performance, learning stability, convergence behavior, and computational trade-offs within the First-Order Model-Agnostic Meta-Learning (FOMAML) framework using an enhanced ResNet-50 backbone. Experiments were conducted on a 3-way classification scenario with 1-shot, 5-shot, and 10-shot configurations. Model evaluation was performed using accuracy, precision, recall, and F1-score, while learning stability was analyzed using standard deviation (Std) and coefficient of variation (CV). The experimental results show that a chili-domain pretrained initialization consistently yields better performance than random initialization. Accuracy reached 95.33%, 95.60%, and 95.94% in the 1-shot, 5-shot, and 10-shot scenarios, respectively an increase of 17.20, 12.61, and 17.49 percentage points compared to random initialization. In terms of stability, the CV values decreased to 1.00%, 0.59%, and 1.03%, compared to 1.10%, 3.66%, and 2.42% with random initialization. These performance improvements were achieved with relatively small differences in training time 0.056 minutes, 0.170 minutes, and 0.876 minutes for the 1-shot, 5-shot, and 10-shot scenarios, respectively. Domain-specific pretrained initialization produces more relevant initial feature representations, thereby improving the effectiveness and stability of FOMAML adaptation while maintaining computational requirements comparable to those of random initialization
Cross-project defect prediction (CPDP) aims to predict software defects in a target project using data from other projects. However, most existing CPDP approaches rely on complex frameworks, such as transfer learning or sophisticated multi-source integration, making them difficult to reproduce and apply in practical software engineering environments. This study proposes a simple yet integrated hybrid preprocessing pipeline consisting of feature normalization, Principal Component Analysis (PCA), SMOTE–Tomek balancing, and decision threshold tuning to improve CPDP performance on the AEEEM dataset. Experiments were conducted under both single-source and multi-source CPDP scenarios using Random Forest (RF) and Support Vector Machine (SVM) classifiers. Performance was evaluated using the F1 Score and the Area Under the Curve (AUC). The experimental results demonstrate that the proposed approach improves prediction performance, particularly under the multi-source CPDP scenario. Compared with the more complex MSCPDP approach, the proposed method achieved a higher F1-score on four out of five target projects and consistently outperformed MSCPDP on all five projects in terms of AUC. Furthermore, the experimental analysis indicates that decision threshold tuning contributed more significantly to performance improvement than class balancing alone. In contrast, the combination of threshold tuning and SMOTE–Tomek yielded the best overall performance. These findings provide empirical evidence that a simple, reproducible preprocessing pipeline can effectively improve CPDP performance without requiring complex learning frameworks.
The rapid development of information and communication technology has driven innovation in various fields, including digital entertainment through video games. This research develops an educational adventure game themed on environmental conservation titled Green Solaris, designed as an interactive medium to raise awareness about environmental issues among children and teenagers. The game adopts a 3D isometric format with a pixelated visual style, focusing on four main themes: Clean Water and Sanitation, Affordable and Clean Energy, Climate Action, and Life on Land. The development process employs the Game Development Life Cycle (GDLC) methodology combined with Agile-SCRUM approach, encompassing six main phases: initiation, pre-production, production, testing, beta, and release. The game features core mechanics including exploration, interaction with NPCs, waste collection and recycling systems, point-based upgrades, and environmental restoration missions across multiple islands. Testing results using Likert scale instruments show that 90% of respondents rated the game as "Very Satisfying" and 10% as "Quite Satisfying" with no negative assessments, indicating the game's potential as an effective educational medium. Green Solaris has been successfully published on the itch.io platform and is accessible to the public, demonstrating that video games can serve as an innovative and engaging alternative learning tool for instilling environmental awareness through creative, interactive, and educational digital experiences.
Burnout has become an increasingly prevalent mental health issue among Generation Z due to the interaction of psychological and behavioral factors in a highly digitalized environment. This study aims to predict burnout risk levels using a multi-class machine learning classification approach. The research follows the Cross-Industry Standard Process for Data Mining (CRISP-DM), encompassing data understanding, preprocessing, modeling, and evaluation. A synthetic dataset containing 10,000 records and 22 psychological, behavioral, and lifestyle attributes was used to classify burnout risk into three categories: low, medium, and high. To address class imbalance and ensure reliable performance estimation, Stratified K-Fold cross-validation was employed. Logistic Regression was implemented as a baseline linear model, while Random Forest represented a non-linear approach. Experimental results demonstrate that Random Forest achieved the best performance, obtaining a macro F1-score of 0.988 and outperforming Logistic Regression in multi-class burnout prediction. Feature importance analysis further revealed that psychological variables, particularly the wellbeing index and anxiety score, contributed more substantially to burnout prediction than behavioral variables such as screen time. These findings indicate that internal psychological conditions are stronger predictors of burnout risk than external digital behaviors. This study provides a comparative evaluation of linear and non-linear machine learning models in an imbalanced multi-class setting and offers an interpretable framework to support data-driven strategies for early burnout detection and mental health intervention.
Achieving successful digital transformation in community-based circular economy organizations requires more than the deployment of digital technologies. Sustainable organizational outcomes depend on the effective interaction of technological capability, organizational readiness, and user engagement. This study evaluates the effectiveness of the IPPEKas information system by integrating the DeLone and McLean Information Systems Success Model, the Human–Organization–Technology Fit (HOT-Fit) framework, and System Usability into a unified analytical model. Data were collected from 100 active IPPEKas users and analyzed using Partial Least Squares Structural Equation Modeling (PLS-SEM). The empirical findings reveal that Information Quality, Organizational Support, and System Usability significantly enhance Net Benefits, whereas System Quality contributes indirectly through its positive influence on System Usability. These findings indicate that organizational value is created primarily through high-quality information and effective user interaction rather than through technical excellence alone. The proposed integrated framework advances information systems success research by combining technological, human, and organizational dimensions into a comprehensive evaluation model. From a practical perspective, the findings provide actionable recommendations for strengthening information governance, encouraging sustained system utilization, and improving organizational support to maximize the long-term value of digital transformation in community-based circular economy organizations.
Social networks are used by individuals not so much for communication as for influencing each other. Social networks are modelled as graphs, and a special Dominating Set is obtained from the resulting graph. A new algorithm has been proposed for the Dominating set, where at least half of the neighbours of each node are elements of the Dominating Set. This algorithm is called as Malatya Positive Influence Algorithm, and its results for some graphs are given in this study.
Multidisipliner çalışma, günümüz bilim dünyasında önemi giderek artan bir konu olmuştur. Özellikle bilgisayar teknolojisindeki gelişmeler, bu bilim dalı ile her disiplini beraber kullanmayı artık bir zorunluluk haline getirmiştir. Bilgisayar mühendisliği akademik çalışmalarında, optimizasyon problemlerine yeni bir bakış açısı getiren meta sezgisel yöntemler; özellikle problemin parametre sayısının ve hacminin büyük olduğu durumlarda ve klasik yollarla problemin çözümünün güç ve çok zaman alması halinde devreye girmektedir. Bu çalışmada literatüre yeni önerilen ve Pi sayısının üretilmesini taklit ederek çalışan yeni bir meta sezgisel yöntem olan Pi Algoritması, hastalık ve gen arasındaki ilişkiyi temsil eden ve Kaggle veri tabanından elde edilmiş bir veri seti üzerine uygulanmıştır. Sonuç olarak; doğruluk için en yüksek 0.96, kesinlik için en yüksek 1, duyarlılık için en yüksek 0.909, F1 skor için en yüksek 0.952 performans değerlerine ulaşılmıştır.
HIV classification using machine learning is often challenged by severe class imbalance, which may reduce predictive reliability and bias classification models toward the majority class. This study aims to compare the effectiveness of two hybrid oversampling techniques, namely SMOTE-DBSCAN and RN-SMOTE, in improving HIV classification performance using an Indonesian HIV dataset. The dataset was obtained from a Non-Governmental Organization (NGO) and initially consisted of 707,389 records with 90 attributes. After data preprocessing, 148,408 validated records with 30 attributes were retained, comprising 135,739 non-reactive and 12,668 reactive HIV cases, indicating a substantial class imbalance. Two feature configurations derived from previous HIV studies were evaluated using SMOTE-DBSCAN and RN-SMOTE. To prevent data leakage, the dataset was partitioned using a stratified 80:20 train-test split before oversampling, while model validation was performed using 10-fold cross-validation on the training data. The balanced datasets were subsequently evaluated using eight classification algorithms. Experimental results demonstrate that RN-SMOTE consistently outperformed SMOTE-DBSCAN across both feature configurations, while Random Forest achieved the best overall predictive performance. Although SMOTE-DBSCAN required less computational time during the balancing process, RN-SMOTE produced more robust classification performance by generating cleaner and more representative minority-class samples. These findings demonstrate the effectiveness of noise-aware hybrid oversampling strategies for improving HIV classification on highly imbalanced datasets and provide practical insights for developing reliable machine learning models to support HIV surveillance and public health decision-making
This paper thoroughly investigates collaborative approaches to enhancing smart vehicles' related cybersecurity through AI-driven threat detection. As connected and automated vehicles (CAVs) become rapidly in demand, new vulnerabilities emerge alongside technological progress. We explored how integration of 5G networks, blockchain system, and quantum computing can address these security related challenges. Our study emphasizes the critical role of intrusion detection systems (IDS), AI-based pattern techniques, and interdisciplinary collaboration across academia, industry, and private sector. We present a roadmap incorporating secure hardware/software stacks and advanced threat intelligence to mitigate cybersecurity threats in autonomous vehicles. We address these challenges, by proposing a multi-layer AI-driven cybersecurity architecture by integrating in-vehicle anomaly detection, cloud-based correlation, and privacy-preserving federated learning. We validated the framework by using a hybrid simulation and edge-device testbed environment. Our results shows improved detection performance (F1-score: 0.97), as well as; enhanced adversarial robustness (89% under FGSM attack), and sub-50 ms real-time response capability while maintaining data privacy through local model training.
This study integrates a hybrid approach to sentiment analysis in the area of medical texts (patient reviews) that merges lexical strategies with deep learning and machine learning models. Different methodologies are used, such as reviews annotated by SenticNet and Text Blob lexicons. and for extracting crucial features in which TF and TF-IDF are utilized. Lastly, classification tasks are performed using machine learning models and deep learning models, and models based on transformers such as Bio-BERT. Performance metrics are utilized to evaluate the effectiveness of this combined methodology. Experimental results demonstrate that hybridization of lexicon and a transformer-based medical learning model produces superior outcomes compared to using each method independently in sentiments finding. Text Blob exhibits impressive performance, achieving 97 percent accuracy with a hybrid of LSTM and CNN, and another medical transformer model is Bio-BERT, on a drug review dataset, with 95 percent accuracy in term frequency and the logistic regression model. TextBlob also attains 94 percent accuracy when paired with term frequency and the LSTM model, and 97 percent accuracy when combined with the Bio-BERT transformer-based model on a dataset sourced from tweets.
This work introduces a dynamic load-balancing algorithm called CHHO (Cuckoo Harris Hawks multiobjective optimization) applied to task scheduling in cloud environment. CHHO is a new hybrid method that combines Cuckoo Search Optimization (CSO) and Harris Hawks Optimization (HHO). This combination uses the strengths of both algorithms to address the complex issues of cloud task scheduling. Specifically, CHHO uses Cuckoo Search Optimization to widen the search area of Harris Hawks Optimization, aiming to improve factors such as cost, response time, and resource use. The CHHO algorithm improves system performance by increasing VM throughput, effectively distributing workloads across VMs and maintaining a balance among task priorities through dynamic adjustments in task waiting times. To test the performance of CHHO, the algorithm is implemented in the CloudSim environment. It is compared with existing load-balancing algorithms on various performance measures. Our simulation results clearly show that CHHO performs better than existing algorithms, providing a strong and efficient solution for load balancing in cloud computing. Introducing CHHO offers a significant advancement in the field, providing a dynamic and adaptable approach that improves cloud task scheduling and enhances the overall efficiency and effectiveness of cloud computing systems.
A knowledge graph is a structured representation of entities and their relationships, often used in biomedical domains to model complex interactions. Graph neural networks (GNNs), which utilize these graphs, are effective for predicting interactions missing in the knowledge graph. However, GNN lacks the ability to incorporate causal reasoning, which is crucial to biomedical applications. Additionally, they limit their ability to generalize to unseen data. In oncology, where treatment regimens are intricate and patient responses are highly variable, predicting adverse drug reactions (ADRs) is particularly difficult. Existing models fail to capture the indirect, high-granularity information needed for accurate ADR prediction. To address these challenges, we propose the Causality and Proximity-based Relational Multihead Attention Model (CPRMAM). This model leverages a knowledge graph of ADR-related cancer case studies and introduces a causal proximity vector to prioritize relevant relationships. By employing an inductive GNN approach, CPRMAM generalizes to unseen data, improving ADR prediction.
In natural language processing with text data, which forms the basis of the studies in the field of artificial intelligence, various studies such as semantics and natural language generation are carried out, especially the solution of classification problems. This study aims to analyze the effect of detected named entities on text classification performance to make the text preprocessing stage more effective. In order to reduce the analysis time and increase the performance, after the classical preprocessing stage, word filtering was performed with Named Entity Recognition according to the thresholds determined in the 5% and 10% ranges. Analysis was performed with various machine learning, deep-learning algorithms, Bidirectional Encoder Representations from Transformers (BERT), and the obtained results are discussed in the last part of the study. In the problem of classifying 50,000 news texts, 93% with a support vector machine (SVM) algorithm in statistical classification with machine learning, 87% with long short-term memory (LSTM), and 83% with BERT success was achieved. In the analyses performed with LSTM and BERT, although the model performances were numerically lower, it was observed that the semantic integrity was stronger in text classification and that the success increased in general after Named Entity Recognition (NER) filtering. Thus, it can be interpreted that the dataset that is passed through the NER filter according to the threshold values positively affects the model's success in terms of time and performance.
Federated learning offers a solution to privacy-related dilemmas of data centralization by maintaining a decentralized architecture, thereby enabling local devices to preserve their data while concurrently exchanging model parameters. Despite its promise, federated learning encounters substantial obstacles concerning data quality, which may arise from inherent biases, the presence of outliers, and the utilization of compromized devices. To mitigate these challenges, we advocate for the implementation of a server-side filtering methodology within federated learning, specifically tailored for regression-related problems. Based on this architecture, local devices train the model on their own data sets and then send the learned parameters to a central server. The server is then tasked with the filtration of erroneous contributions, thereby enhancing the overall accuracy of the model. This methodology is substantiated through the application of the mean squared error (MSE) metric, a widely recognized standard within regression analysis, thereby augmenting both the efficiency and dependability of the learning process while safeguarding user privacy, an essential component of federated learning.
Air pollution remains one of the most critical environmental challenges of the 21st century, with severe implications for public health and urban sustainability. To address the need for accurate and timely air quality forecasting, this study introduces DeepBiGRULSTM, a hybrid deep-learning framework that memory (LSTM) networks. The model leverages both real-time Nowcast and raw concentration values to enhance the reliability of short-term predictions. Using a comprehensive historical dataset from the U.S. Dhaka consulate, incorporating pollutant and meteorological variables, we benchmarked the proposed model against established architectures. such as GRU, LSTM, and Temporal BiGRULSTM consistently outperforms the baselines, achieving the lowest errors with an RMSE of 0.4976 on the training set and 0.7548 on the test set. A key contribution of this work is its integration into a mobile application, providing real-time air quality updates and early warnings across Bangladesh. Beyond forecasting, the system supports health advisory services and can inform public policy on environmental management. Overall, this research delivers a practical and high-precision framework for air quality monitoring, with direct relevance to sustainable urban development and public health protection.
